Author: Su Qingtao
Although Mobileye has been willing to promote “software and hardware decoupling”, because car companies currently do not have the ability to handle algorithms, Mobileye has to return to the old path of software and hardware integration.
With the increasing strength of software capabilities, in cooperation with car companies, does Nvidia have the opportunity to change “Nvidia Inside” to “Nvidia Outside”?
In the end, Mobileye and Nvidia, Qualcomm, and Horizon will converge on the commercial model, that car companies can buy chips alone, or chips + algorithms.
Car companies that give up their self-developed autonomous driving chips are likely to choose different chip suppliers and solutions for different car models based on product strength and cost.
The material for this article has been lying in my draft document for a year, but unfortunately, I haven’t had time to sort it out until recently.
I initially planned to use the title “Nvidia is turning soft”. In fact, many actions taken by Nvidia, Huawei, Horizon and many other chip manufacturers in the past year echoed the theme of “Nvidia is turning soft.” Recently, however, we learned that a certain Asian car company’s cooperation with Mobileye based on EyeQ 5 did not proceed in the direction expected by both parties, hence the first half of the title.
I’m not using one or two cases to “force relationships.” In fact, as early as the first half of last year, the founder of a Chinese startup who had contact with both Mobileye and Nvidia said, “Mobileye is becoming more and more like Nvidia, and Nvidia is becoming more and more like Mobileye.”
The veteran said that Mobileye “is becoming more and more like Nvidia,” referring to its need to become “open” in the face of new competitive situations, while his vision of Nvidia “becoming more and more like Mobileye” confirms Horizon’s founder, Yu Kai’s observations from as far back as 2017: “We see a trend where semiconductor manufacturers are not just doing hardware, but are increasingly moving upstream to build software architectures.”
Mobileye: I can be open, but can you reach my level of ability?
Mobileye’s “closed” (highly coupled software and hardware) business model has always been criticized by car companies, and Mobileye itself is aware that a closed mode will continue to face challenges from customers who intend to develop their own perception algorithms.
In fact, Mobileye also has an “open mind,” but for various reasons, it has not been implemented in practice.As early as May 17, 2016, in the joint press release of Mobileye and STMicroelectronics announcing the development of EyeQ5 (of which the manufacturing was later outsourced to TSMC), it was mentioned that EyeQ5 would be developed as an open platform, with two sales models: “chip + algorithm” and “chip only,” the latter allowing customers to write their own algorithms based on their specific needs.
However, at that time, except for Tesla, no other automakers had a strong demand for “hardware-software decoupling,” so most people didn’t pay much attention to Mobileye’s “open” plan. As a result, the plan did not spark much discussion at the time, and people even forgot about it when criticizing Mobileye for its “closed” approach.
Mobileye CEO Shashua only mentioned in 2018 that opening up a SoC that was originally closed required a lot of resources, and Mobileye simply did not have the capacity until August 2017, when Intel completed its acquisition of Mobileye and dispatched 200 engineers specifically to develop the EyeQ5 open architecture.
At CES in January 2018, Amnon Shashua reiterated the almost-forgotten plan, saying that Mobileye would provide a machine learning-based SDK (software development kit) for EyeQ5 users to write their own algorithms/programs.
However, Mobileye’s “openness” still did not attract much attention, and was soon forgotten.
In October 2018, in an interview with Bloomberg, Amnon Shashua once again systematically explained Mobileye’s “reform and opening-up” plan starting with EyeQ5.
In November, in a report by EE Times, Amnon Shashua emphasized: “Mobileye is unique because we are the only company pushing for an ‘open’ strategy, leaving room for cooperation with first-tier suppliers and automakers.” This statement completely disregarded the fact that Nvidia had already implemented an open strategy and Chinese chipmaker Horizon had implemented an open business model from the beginning.However, for automakers who have been struggling with “unchangeable algorithms” but have employed Mobileye’s self-driving chips, Mobileye’s “reforms and opening up” undoubtedly provide huge benefits. In theory, automakers will welcome Mobileye’s move with great fanfare.
Under Mobileye’s original plan, its close partner BMW would be the first customer for EyeQ 5. However, in the second half of 2020, an Asia-based automaker announced it would use EyeQ 5 ahead of BMW. According to information disclosed by both parties during media interviews at the time, in this collaboration, Mobileye officially began to try “software and hardware decoupling”, that is, allowing automakers to write their own perception algorithms.
Moreover, I also learned from industry insiders that Mobileye has allowed the automaker to add “additional” sensors to run shadow mode on the car in this collaboration, which was strictly forbidden for Tesla and other Mobileye EyeQ 3 and EyeQ 4 customers before. This time, Mobileye’s openness seems to be “very sincere.”
However, the words of executives from both sides in media interviews about how to open EyeQ 5 and to what extent are worth careful consideration. The Asia-based automaker’s executive said, “The core of software and hardware decoupling is that software collaboration will eventually return to chip collaboration, which is one of the reasons we directly collaborate with Mobileye. We value Mobileye’s leading global vision technology, while we are not very good at this area.”
The former seems to mean “Mobileye only sells chips, and we write the algorithms ourselves”, while the latter seems to say “our involvement in the algorithms is not very high.” Erez Dagan, Vice President of Product and Strategy Execution at Mobileye, said, “The industry trend we see is that, except for companies with sufficient funds to bear risks and failures, most companies have realized that developing the entire stack solution is a high-cost, high-risk game. Software and hardware decoupling will not be the optimal solution. A tightly integrated system is bound to be more efficient and performance will continue to improve. We are very optimistic that the integration of the largest and most famous host manufacturers will adopt a flexible and tested full-stack solution.”
It can be understood that although Mobileye has made an open gesture, they are still unwilling to completely open up and hope to retain a certain degree of “software and hardware coupling”.At the time, the author’s judgment was: to cope with the trend of software and hardware decoupling, Mobileye would definitely fulfill its “open” commitment, but the degree of openness and the extent to which customers can participate still require a long-term game between the two sides.
Recently, the author received information from insiders that the perception and decision-making algorithms in this project are provided by Mobileye, and the OEM only writes the control algorithm. The reason is that the project is too tight on schedule, and “customized development” is not allowed at all.
This means that although Mobileye is already willing to promote “software and hardware decoupling”, due to the fact that the OEM currently does not have the ability to handle algorithms, Mobileye has “to” return to the old road of software and hardware integration.
NVIDIA: I cannot give up those customers with “poor algorithm capabilities”
In fact, whether to decouple software and hardware depends not only on the attitude of chip manufacturers, but also on the software capabilities of the OEMs, which is also a very important factor. This can also be seen from NVIDIA’s list of mass-produced customers for its “Drive PX 2” platform, which is geared towards autonomous driving.
The real mass production customer of the first autonomous driving computing platform “Drive PX 2” was Tesla; the first mass production customer of the second-generation autonomous driving chip “Xavier” worldwide was XPeng, and the second was Toyota; and the first mass production customers in the world for the third-generation autonomous driving chip “Orin” were Ideal, XPeng, NIO and other Chinese automakers.
Although the product performance is powerful, the customer base is far less than Mobileye’s EyeQ series. Why is that? Is it just because the price is too high?
In addition to providing higher computing power, the biggest advantage of NVIDIA compared to Mobileye is its “openness”, but this also means a high usage threshold. OEMs must handle the algorithms themselves (or cooperate with Tier 1 to handle the algorithms). However, the current situation is that at this stage, the vast majority of OEMs do not have the ability, and even the confidence, to handle the algorithms themselves.
The first batch of mass production customers for each generation of NVIDIA’s products are limited to those OEMs that have the ability or confidence to handle perception algorithms.
Developing perception algorithms in-house is extremely difficult.
Even Tesla, after switching from Mobileye’s EyeQ 3 to NVIDIA’s Drive PX 2 and using its own algorithms at the end of 2016, once experienced the problem of AEB function downgrading – as long as the vehicle speed is higher than 28 miles/hour, AEB cannot come into effect, while in the Autopilot 1.0 era, as long as the vehicle speed does not exceed 85 miles/hour, AEB can come into effect. And the reason is that at that time, Tesla’s algorithm was not good enough.After experiencing the difficulty of the challenge with the use of Xavier and self-developed algorithms, XPeng, just like Tesla, polished his self-developed algorithms over several months and made them mature. However, most car companies cannot compete with Tesla or XPeng in terms of software algorithm capability, so although they know that NVIDIA chips are powerful, they dare not purchase them in large quantities. Tier 1 companies like Bosch also have the same problem.
As early as March 16, 2017, Bosch announced a strategic cooperation with NVIDIA, which had just entered the autonomous driving industry at that time: according to the agreement, NVIDIA chips would be applied to Bosch’s design of autonomous driving domain controllers in the next few years. At that time, NVIDIA wanted to take advantage of Bosch’s influence.
A month later, Bosch partnered with Mercedes-Benz to establish the Autonomous Driving Alliance. It seems that Mercedes-Benz could become NVIDIA’s first indirect customer through Bosch.
In July 2018, NVIDIA officially announced its three-party cooperation with Bosch and Mercedes-Benz. The Robotaxi trial operation vehicles launched by Mercedes-Benz and Bosch in Silicon Valley in early 2020 were equipped with NVIDIA’s Drive Pegasus platform.
Bosch hopes to make this alliance an open platform. In early 2019, when talking about cooperation with Mercedes-Benz, Bosch CEO Volkmar Denner said, “We are open to other partners and welcome negotiations.”
However, neither Bosch nor Mercedes-Benz can handle perception algorithms. A friend in Bosch’s German headquarters said: “German car companies and Tier 1 companies are too focused on lean production in the mechanical aspect and ignore the rapid replacement of software application environments, which leads to outsourcing a large number of software businesses. Moreover, the development process and thinking mode of companies branded with the label of mechanical and electronic development in Germany are not suitable for software development.”
NVIDIA certainly does not want to “let go” of most car companies, which account for the majority of the market. Letting them go is equivalent to giving up the largest market. Therefore, NVIDIA must find a way to make these algorithmically weaker car companies dare to purchase their chips in large quantities.
NVIDIA’s method is simple: to provide these car companies with a complete solution of chips + algorithms, which is what Mobileye has been doing all along.
The result is what we see today. At the end of June 2020, Mercedes-Benz reached a strategic cooperation with NVIDIA: from 2024, NVIDIA’s autonomous driving chips and algorithms will be installed on all Mercedes-Benz models, and the algorithms will be jointly developed by the two parties, mainly developed by NVIDIA. This is NVIDIA’s first public appearance as an “autonomous driving algorithm supplier”.In early 2021, when asked by the media whether NVIDIA would be a hardware supplier, a software supplier, or both, Huang Renxun answered, “If Robotaxi wants to develop software independently and operate the business by themselves, we’ll sell chips to them and provide them with toolchains. Some customers hope that we can provide a full-stack solution, including software, so that they can have this capability before 2022.”
A source from a domestic new energy vehicle company with a strong interest in NVIDIA reveals that NVIDIA will also provide the algorithm developed in cooperation with Mercedes-Benz to other automakers, and Mercedes-Benz will receive a certain percentage of the revenue.
When “automakers develop their own chips” becomes a trend
Before 2019, NVIDIA was far ahead in high computing power autonomous driving chips. However, since then, Huawei, Qualcomm, and Waymo, which are all moving toward tier-one transformation, have all started to compete with NVIDIA for market share. However, over a longer period of time, both NVIDIA, Huawei, Qualcomm, and Waymo will have to face challenges from automakers developing their own chips.
Tesla was the first customer of NVIDIA PX2. However, in the second half of 2017, Tesla officially admitted that it was developing autonomous driving chips on its own. In fact, Tesla had already been preparing for the development of their own chips before partnering with NVIDIA. Since April 2019, Tesla’s new cars have all been equipped with FSD chips developed in-house, and NVIDIA has since withdrawn.
Tesla’s self-developed chips have a strong demonstration effect on other automakers.
As early as before 2017, Chinese new energy car maker Zero Run began to develop autonomous driving chips with Dahua Weishi. Although the outside world was not optimistic, in October 2020, their first self-developed autonomous driving chip, Lingxin 01, was officially released, and this chip will be installed in their next car model, the C11; subsequently, Lingxin 01 will also be installed in the existing car models, S01 and T03, in the new year’s models.
In 2020, several automakers announced plans to develop their own autonomous driving chips.
In April, Toyota and Denso invested 460,000 USD to establish a joint venture, MIRISE Technologies, to produce “car semiconductors,” with a share ratio of 49:51%. Although the registered capital was small, the goal was to conduct research and development of autonomous driving chips.
In May, BAIC and UK chip design company Imagination established a joint venture, and in October, Geely’s subsidiary, Yikatong, and chip design company ARM established the design company, Horizon Robotics. Both joint ventures include the development of self-developed autonomous driving chips as part of their mission.
Among China’s new automakers, NIO explicitly mentioned the development of their own chips in the first half of 2020, and there were rumors that NIO had established a chip team named Smart HW (Hardware).In early 2021, during a conversation with Professor Zhao Fuquan from Tsinghua University, Li Bin mentioned that the cost of self-developed chips is too high, and emphasized that “the premise of self-developed chips is that the algorithm has been finalized and the coupling degree between the algorithm and the purchased chip is poor; if the algorithm is uncertain at present or the existing chip is already well-functioning, there is no need to develop dedicated chips.” In response, a high-level executive of a chip manufacturer explained that NIO did establish a chip team and has been developing self-driving chips. Li Bin’s emphasis on the difficulty of self-development in interviews was only to “appease suppliers” because the development cycle of chips takes at least three years.
By the end of January during a talk with China International Capital Corporation Limited, Li Bin said that “people who are serious about software must also be serious about hardware. The boundary between software and hardware is not clear, and software often solidifies into hardware. This trend is inevitable. As for how to do it, whether to self-develop or cooperate with others, it’s not necessarily fixed.”
Among domestic car companies, XPeng Motors has the highest investment in self-driving technology. Previously asked by the media if XPeng Motors would self-develop chips, the head of XPeng Motors’ self-driving department, Wu Xinzhou, replied, “We won’t say we will definitely not develop chips. I personally have a background in chips, so our understanding of chips is okay.”
During the Q4 earnings call in 2020, XPeng Motors CEO He XPeng stated that “in 2021, we will increase research and development investment, including closely-related hardware for autonomous driving.”
By April 2021, 36Kr reported that XPeng Motors had indeed entered the chip field. Benny Katibian, Chief Operating Officer of XPeng North America, and Xia Heng, co-president of XPeng Motors, are in charge of the chip project in North America and China respectively. The report stated that “if the progress is smooth, XPeng’s chip is expected to be taped out by the end of this year or early next year.”
Ideanomics also discussed plans to develop self-driving chips in 2020, but later became more cautious as they considered that the premise of developing good chips was having a highly-developed algorithm. However, in April 2021, they announced plans to make autonomous driving features standard in new cars. Considering that both NIO and XPeng Motors have already taken action, Ideanomics will most likely also develop their own chips.
In early May, Herbert Diess, the CEO of Volkswagen, announced that Volkswagen will independently design and develop high-performance chips for its autonomous driving cars, in order to break away from the constraints of chip manufacturers’ product iteration speed. As Wu Xinzhou said in an interview last year, “Our Xavier is not yet the most powerful chip, but we actually have no better choice. The higher performance (chip) will not be available for at least another two years. However, it is entirely possible that in two years, competitors may reach 700 Tops.”## Hardware optimization for smart cars
Hardware for smart cars needs a lot of performance optimization for specific application and algorithms. Separation of software and hardware can result in suboptimal performance. Maximizing the benefits of software differentiation requires the use of dedicated chips.
The author once asked the CMO of an autonomous driving chip manufacturer if they could provide a chip like Tesla’s Hardware 4.0, as a central supercomputer, which is an AI chip that can handle all domain calculations. Their answer was, “We can do it in terms of capability, but we won’t do it commercially because Tesla is a host manufacturer that only needs to consider its own requirements. As a chip manufacturer, we need to balance the needs of all customers, so the chip needs to be more universal.” It seems that this chip that supports central supercomputing can only be achieved through self-developed chips by car companies.
In addition, many car companies plan to self-develop automatic driving operating systems to better support automatic driving applications. However, car companies or autonomous driving companies that have already self-developed successful automatic driving operating systems, such as Tesla, Waymo, Mobileye, Huawei, and Apple, all have one thing in common: self-developed ASIC chips. This may mean that car companies that plan to self-develop automatic driving operating systems need to also develop chips.
At the “feasibility” level, car companies that develop ASIC chips are more likely to achieve stronger computing power. For example, Nvidia’s Xavier has 90 transistors, Tesla’s FSD chip has 6 billion transistors, and the latter is 2.4 times faster than the former in terms of computing power. There are two possible reasons for this: 1. Tesla’s chip designer Peter Bannon and his team are stronger than Nvidia’s designers; 2. Tesla’s software-hardware combination methodology (designing hardware based on software algorithm operating requirements) is better.
Furthermore, when car companies make ASIC chips, they can cut out a lot of generic interfaces that are not related to specific requirements. Therefore, the process is simpler, the design speed is faster, and the cost is lower.
Referring to Apple’s case in PC chips, it is clear that the performance improvement of self-developed chips by car companies will be very fast. In the early days of self-developed Apple chips, their chip performance was significantly different from that of traditional chip manufacturers’ SoC performance. However, through system-level optimization, the final experience difference was not significant. With years of experience accumulation and chip iteration, Apple’s self-developed chips have become powerful enough to replace Intel’s CPU.Some industry insiders believe that the production of general-purpose chips needs to reach a volume of one million chips in order to break even. According to consulting firm Fraux, assuming an OEM spends $150 million on design costs for its self-driving chips, the costs could be recouped after four years with 400,000 units produced per year without a change in component prices. NIO conducted a test of its self-developed self-driving chips and found that the cost was about $200 million.
Therefore, it has almost become an irreversible trend for top automakers to develop their own chips. This is not a good sign for chip manufacturers such as NVIDIA.
According to The New York Times, NVIDIA attempted to acquire Zoox at the end of 2019 or early 2020, but talks fell through. If the acquisition had been completed, NVIDIA could have become a Robotaxi operator like Mobileye. Unfortunately, this did not happen.
Of course, establishing deep cooperation with Daimler, which has an annual sales volume of more than 2.3 million vehicles, is also a good choice. Through this cooperation, NVIDIA can provide algorithms on top of chips to form a complete solution, and can offset the impact of automakers’ self-developed chips by providing this solution to other automakers under agreement with Daimler.
Shortly after the agreement was announced, Huang Renxun said on an investor conference call that this was “the biggest single business model shift” in the company’s history. Danny Shapiro, Senior Director of NVIDIA’s Automotive Business, also said, “Don’t think of this partnership as a product, but as a strategic and architectural change for the entire product line in the future.”
NVIDIA’s Accumulation of Software Algorithm Capabilities
Of course, NVIDIA’s offer of algorithms to automakers is not a hasty response to the rapidly changing competitive environment, but rather a “natural fit” after years of accumulation.
In 2007, NVIDIA officially launched CUDA, a GPU unified computing architecture platform, which is equivalent to packaging complex graphics card programming into a simple interface, benefiting a large number of programmers. NVIDIA has developed and accumulated a large number of algorithms and software for different fields based on CUDA, allowing countless developers to run a basic deep learning model without writing a line of code, and upgrade and optimize their software stacks by standing on the shoulders of giants.
Due to the huge number of developers, NVIDIA DRIVE has become an industry standard for the development of autonomous driving cars, adopted widely by automakers, truck makers, Robotaxi companies, software companies, and universities. The CUDA ecosystem has become NVIDIA’s most important moat.In early 2017 CES, NVIDIA released its autonomous driving platform NVIDIA DRIVE, which is paired with its self-developed software architectures Drive AV and Drive IX. The so-called Drive AV integrates deep neural networks, which can perceive raw data, understand the surrounding environment of the vehicle, and predict the actions of other road participants. As for Drive IX, it is an intelligent cockpit software that can run on Xavier, and it provides tools for supervising systems and various user interfaces for drivers.
In March 2018, NVIDIA released an autonomous driving simulation system, Drive Constellation. The project that Volvo cooperates with NVIDIA uses Drive Constellation simulation platform for testing.
During CES in January 2019, NVIDIA demonstrated its high-precision localization solution DRIVE Localization (software module). In addition, NVIDIA is also planning a high-precision map crowdsourcing scheme, NVIDIA MapWorks.
At the GTC conference in December 2019, NVIDIA did not launch any performance-exploding chips. Instead, more new products focused on the software level. Products such as AP2X 9.0, DRIVE Constellation, and Safety Force Field driving strategies were all software products, indicating that the company has comprehensively turned to software systems in the field of autonomous driving.
Compared with the previous versions, AP2X 9.0 has added a “blindness” warning function for cameras. That is, NVIDIA’s self-developed deep neural network (DNN) ClearSightNet evaluates the visibility of the camera to determine the root cause of obstruction, obstacles, and visibility reductions. In this way, the perception system can detect invalid data as early as possible in the processing pipeline before the data is processed by downstream modules. With this function, when the perception data finally reaches the decision-making end, the vehicle can choose not to activate the autonomous driving function, and remind the user to clean the camera lens or windshield, or use ClearSightNet output to notify the user of the calculation results of the camera perception credibility. Moreover, it can maximize security when the camera goes “blind.”
Also in December 2019, NVIDIA open-sourced the development of deep neural networks for autonomous driving car developers. Through this AI tool, developers in the NVIDIA ecosystem can freely expand and customize the model, thereby improving the robustness and capabilities of their autonomous driving systems.
NVIDIA has the most stable toolchain in the industry.Last year, I asked a well-known chip manufacturer how many people are needed to develop an autonomous driving chip. I questioned, “Why did Tesla produce its chip with fewer than 100 employees while Horizon Robotics needed at least 500 and NVIDIA’s Xavier, as rumored, required 2,000? Where does the difference lie?” The wise man replied, “Tesla’s chip is only for internal use, so it is simple to produce and does not require many workers. In contrast, Horizon Robotics is more open and requires many workers to create tools. NVIDIA’s products are even more universal than Horizon Robotics, so its tool requirements are even higher, so NVIDIA’s investment of over 2,000 people in Xavier is mostly used for the creation of tools.”
Last September, when Ideal Auto signed with NVIDIA, Ideal’s CTO Wang Kai revealed that they chose NVIDIA because it has a stable toolchain and rich software ecology. “The benefit of a mature toolchain is that if something goes wrong, you can quickly determine if it is caused by an unstable toolchain, unstable hardware, or an unstable program.”
Behind the toolchain is software capability.
At the beginning of 2021, after the term “Huang’s Law” (which refers to NVIDIA’s AI chip iteration speed surpassing Moore’s Law) became popular, NVIDIA expressed that it values chip architecture optimization over the advancement of semiconductor manufacturing technology. “Behind the chip architecture lies software capability.”
Despite chip sales still being NVIDIA’s primary income source, the number of software employees in its team exceeds that of hardware employees. In the first half of 2020, executives of new domestic car companies revealed that NVIDIA had over 1,000 autonomous driving algorithm engineers.
In fact, as early as the Q3 earnings call in October 2019, Huang Renxun (NVIDIA CEO) announced, “NVIDIA has become a software company.”
With the increasing strength of software capability, NVIDIA has the opportunity to turn “Nvidia Inside” into “Nvidia Outside” in collaboration with car companies.
At the beginning of 2021, Huang Renxun said in a media interview, “In the past, the day a car was picked up was its ‘peak’. However, a software-defined car will be the exact opposite. The day the car is picked up will be its ‘trough’. But after that, it will be wonderful.” He also said, “New cars sold at cost in the next four years will no longer be a ‘utopia’ because profits will come from software. If a car company produces 10 million cars per year and sells them at cost, it is possible to generate a profit of \$5,000 from software, resulting in a total profit of \$500 billion for that year.”
Since Huang Renxun is fully aware of the importance of software algorithms in the automotive industry’s value chain, wouldn’t he plan to take a slice of the pie?>When asked “Who will dominate the software in cars?”, Huang Renxun said that the difference between cars and mobile phones is that the onboard software is customized for vehicles. “The idea that the technology industry will control everything in the car is unfounded and will not happen. The reason is simple: automakers will become fleet managers and service providers, rather than just mechanical manufacturers. For a long time, automakers will largely have the dominant power of software, which is fundamentally different from the mobile phone industry,” he said.
It appears that Nvidia does not intend to “control” automakers even though they have the ability to provide algorithms. We don’t know if this is Nvidia’s genuine thought or just an attempt to appease its customers. Nonetheless, compared to some tech companies that attempt to make automakers work for them, this philosophy is undoubtedly more reassuring for automakers.
Some additions to this article:
- Chip manufacturers provide software algorithms based on hardware capabilities, not just Nvidia, but also Huawei, Qualcomm, and Horizon.
We are all familiar with Huawei’s “full-stack” plan. Qualcomm, after launching Snapdragon Ride in 2020, has also begun to integrate software ecology, including joint ventures with Veoneer to create ADAS and autonomous driving full-stack software, which many people may not have noticed.
Horizon is even more interesting. The founding team of Horizon is mainly from an algorithm background, and while they were making chips, they did not emphasize the value of algorithms.
In November 2017, Horizon’s founder, Yu Kai, said at the Future Transportation Forum in Hangzhou, “We see a trend in the direction of autonomous driving. Semiconductor manufacturers not only do hardware but also increasingly build software architecture; on the other hand, traditional software players who have only done software in the past are now moving toward a direction of software-hardware integration, such as Google, which is also making chips. We believe that to truly solve the landing of AI applications, we must adopt a software-hardware integration approach.”
After making the chip, Horizon’s identity as a software-hardware integrated solution provider has become increasingly clear.
At the Shanghai Auto Show, Horizon released the Horizon Matrix Pilot based on the Journey 3. From the name of the Pilot, it can be seen that this solution rivals popular navigation-assisted driving features such as Tesla’s NOA, XPeng’s NGP, and NIO’s NOP.
In the early days of its founding, Horizon’s positioning was to become the “Intel Inside” of the AI era, but now they emphasize “AI on Horizon”.
The market size of being a complete solution provider is obviously larger than that of being a chip supplier.2. Regarding the topic of self-developed chips by car manufacturers, Yu Kai responded in an interview with Automotive Industry Review in March by stating that:
“It’s hard to say about this thing. Some phone manufacturers have also tried to make phone chips and later gave up. Many computer companies also tried to make their own chips in the early days, either failing or giving up, such as SUN, IBM, and PC-era Apple. We feel that car manufacturers find it difficult to make chips because firstly, the sales volume of a car factory is not so large, generally ranging from a few hundred thousand to a million, making it difficult to dilute the chip R&D cost; and secondly, making chips by oneself will isolate oneself from the open software ecology, because it is difficult to find software talent familiar with closed chip platforms in the industry.”
“As for where the focus of competition is for most engine manufacturers? First of all, you need to make good cars, sell them well, establish a good brand, and maintain customers. These things are all difficult, and your opponents are very strong, so you must be fully focused on them. Only then will you have the spare energy to develop chips. Honestly, not everyone can be like Steve Jobs.”
- Regarding the development of algorithms by chip manufacturers for car manufacturers, industry insiders commented as follows:
“Who will take the lead, car manufacturers or chip manufacturers? Can we consider chip manufacturers as suppliers to car manufacturers? Not quite, because the core technology is all in the hands of chip manufacturers. So, should we consider car manufacturers as peripheral suppliers to chip manufacturers? This is also not quite feasible, as car manufacturers certainly won’t obediently be under someone else’s control. The simplest question is: should car manufacturers actively cooperate with the iteration speed of chip manufacturers or should chip manufacturers actively cooperate with the development speed of car manufacturers? This is two completely different schemes.
Moreover, if two competing car models use the same solution from the same chip manufacturer, the core technology will be similar, but how should they compete? How should consumers choose? This is different from things like chips and lenses in phones – phone manufacturers need to adapt deeply themselves to have differentiated competitiveness. Can car manufacturers do the same?
Furthermore, this cooperation essentially transfers the task of intelligence from car manufacturers to chip manufacturers. One chip manufacturer may correspond to dozens of different car models from multiple car manufacturers. Can they really meet the differentiated competition needs? Like in schools, one teacher can teach a few students well based on their individual needs, but if one teacher corresponds to 30 or 40 students, it will be very difficult. With too many students, they can only have one-size-fits-all education. Can chip manufacturers do this? This requires chip manufacturers to have strengths and energy several times stronger than Tesla and new forces, and can the communication between chip manufacturers and car manufacturers be as smooth as between two departments?
Therefore, after several rounds of negotiations, the final strategy for chip manufacturers may be to only deeply bind with a few car manufacturers, and only sell chips to most car manufacturers.”
- My colleague, Professor Sun Li’s view is:Those car companies that have given up the development of self-driving chips may choose different chip suppliers and solutions for different car models based on their product strength and cost. For instance, Great Wall Motors has selected both Huawei and Qualcomm as suppliers for different cars. However, they may also choose a primary strategic partner in order to develop deep customizations and differentiation.
Since the future of automobiles lies in software and hardware integration, it is crucial to see who can do it well, and it depends on who has the best strategic alliance.
To create a good alliance, both parties need to have strong capabilities and collaborative abilities, and if one party “lags behind,” it may not work. For example, the cooperation relationship between Xiaomi and Qualcomm in the mobile phone industry has had a significant impact on Xiaomi’s growth, enabling them to frequently obtain “first launch” and be one step ahead in optimization.
Regarding the strategic cooperation between Mercedes-Benz and NVIDIA, from publicly available information, this cooperation is exclusive to Mercedes-Benz, meaning NVIDIA will not work with other chip factories. But is this exclusivity also for NVIDIA? Will NVIDIA have another comprehensive strategic partner who is involved in research and development and sharing of profits, like Mercedes-Benz? Perhaps Mercedes-Benz has also put restrictions on NVIDIA, saying that “it is okay to work with other car companies, but the depth of collaboration cannot exceed mine”?
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.