Source: Far River Research Institute Author: Xiong Yuxiang
Cover/Figureworm Creative Editor: Zhu Hao
In March 2017, Intel acquired Israeli self-driving chip company Mobileye for $15.3 billion, and then-Intel CEO Kozanich had high hopes for the acquisition, calling it "the combination of brain and eye of self-driving cars." Mobileye founder Amanon Shashua believes that with Intel, an Israeli company will be more capable of changing the world. Reuters said in the headline of the report that the acquisition "shook the autonomous driving industry".
Five years later, when the competition for autonomous driving is in full swing, Intel made a sound in March this year, and the spin-off of Mobileye is underway for independent listing. Although the rumor market has given a $50 billion valuation, the comments about Mobileye have not given much face, and some voices have called it "high cash-out". Intel and Mobileye's imaginary "strong alliance, defend the throne" situation did not appear.
When the acquisition occurred, Mobileye was the number one self-driving chip. Shashua, dean of computer science at the Hebrew University in Israel, solidified the computer vision software algorithm he mastered into the chip, encapsulated a set of assisted driving solutions, and helped car companies achieve the assisted driving function at a lower cost and development resources. When the industry's goal moved from assisted driving to autonomous driving, Mobileye rose to the top of the autopilot chip.
At its peak, Mobileye's market share exceeded 70%, and the automotive industry called it "the main engine factory of flowing water, the iron mobileye".
The first on the right is Shashua
In recent years, Intel and Mobileye's competitors have successively come to the front, not only the chip companies Nvidia and Qualcomm, which also originated in the United States, but also Huawei, Horizon, and Black Sesame, which are rising under pressure on the other side of the ocean. The rising crowd has impacted Mobileye's position in different ways, some are known for product performance, and some are attracting customers with warm service.
Car companies continue to "defect" from Mobileye's camp. Last year, Mobileye's seed customer, BMW, with whom it has been working for fifteen years, also chose to "flee."
The self-driving chip leader is facing a "peak crisis": the dominance of the old days brought Mobileye the high valuation it has today, but its position for the future is constantly disintegrating.
Hejiu bifen does not constitute a full explanation of Mobileye's situation, and everything was buried as early as 2016 and even earlier.
01/Reverse package: Open the black box of the autopilot chip
In April 2016, a Tesla Model S crashed into a truck in the Middle of a Road in the United States, killing the driver on the spot. The Tesla turned on the driver assistance function Autopilot, and the core of the latter's perception system was a Mobileye EyeQ3 chip.
Tesla's first-generation driver assistance system board
After the Shashua accident, he publicly accused Tesla of being too aggressive, "challenging the limits of safety" and "harming the interests of the company and even harming the industry." Considering the industry's rules for automotive suppliers not to say bad things about car companies' customers, Shashua's statement reflects the fact that the two sides have a deep grudge.
In fact, Shashua angered Tesla not only because of the loss of life, but also because Tesla had long been plotting a betrayal: he mastered the core technology of autonomous driving and was ready to kick away Mobileye when the time was right. Mobileye has threatened Tesla with price increases, but Musk has long figured out that "our rupture is inevitable."
In 2016, Musk developed a grand blueprint for Tesla 2.0, which included a key KPI, "to develop self-driving technology that is 10 times safer than human manual driving." The problem was then exposed: under the demand of leapfrog development, the traditional packaging model that Mobileye is good at for the automotive industry has gradually failed.
For a long time, packaging was the basic way the automotive industry was organized. Tier-2, Tier-1, car companies form a pyramid industrial chain, car companies at the top responsible for product function definition and technology integration, Tier-2 and Tier-1 will be their master of the technology packaged, packaged into a car company only need a small amount of development, testing, verification can be used to use the module.
In an automotive industry with a high degree of division of labor and packaging everywhere, the mastery of technology by car companies is often concentrated in the parts they think are the most critical: such as engines, gearboxes, chassis (in fact, there are countless packaging solutions for traditional automobiles that car companies can use), and a large number of research and development work is actually delegated to the hands of suppliers.
This extreme division of labor has led to the optimization of efficiency, the reduction of costs and the increase of production, and the globalization of Dongfeng, global automobile sales have increased from 58 million in 2000 to nearly 95 million in 2016. Similarly, with mobileye's packaged EyeQ chip, the automotive industry has achieved the initial popularization of assisted driving.
However, encapsulation is not without cost.
The supplier "out of good intentions" provides a packaged solution for downstream customers, which means that car companies are often faced with a black box that knows but does not know why, and can only call and cannot penetrate, and then lose control of the underlying technology.
Suppliers are not only often the source of new technologies, but also the ceiling.
Mobileye's EyeQ3 was already the most advanced and reliable chip Tesla could find at the time, but the demand for autonomous driving far exceeded its cap. And, in order to defend its intellectual property and business model, Mobileye "politely rejected" Tesla's request to develop its own self-driving technology on its chips and algorithms.
Such a huge black box blocking the way completely annoyed Musk, who revered the first principles. In 2016, Musk established Tesla's machine vision team to develop perception algorithms from scratch; and then dug up the "silicon immortal" Jim Keller from AMD, developed his own chip, and broke through the bottleneck of automatic driving in two ways [1].
Tesla's approach, in the following years, like ripples, has successively radiated new car-making forces, independent car companies, and foreign-funded car companies. Looking back, the fatal car accident in April 2016 was actually the fuse of a wave of automotive automotive anti-encapsulation in car companies.
However, at that time, the small-scale Tesla gave extremely limited orders, and Mobileye did not think that there was anything wrong with the model of "doing more and customers doing less", after all, not every car company was Tesla, and had the strength and courage to develop self-developed autopilot chips and algorithms.
The results seem to prove it: between 2016 and 2020, Mobileye's annual chip shipments more than tripled from 6 million to 19.3 million.
However, the Israelite Shashua probably did not hear an Oriental proverb: the spark of a star can burn the plains.
02/Soft-hard decoupling: from independent kingdoms to open platforms
In January 2021, SAIC's high-end brand Zhiji New Car has just debuted, and it wrote an important autonomous driving chip partner on PPT, NVIDIA. What is less well known is that SAIC is a key partner of Mobileye in China, launching an EyeQ-equipped model in 2016 and discussing strengthening cooperation on autonomous driving in 2020.
Why did SAIC transfer love in only 1 year? According to industry popular saying, the combination with Mobileye is "soulless".
Car companies are not pure assembly plants, but will selectively invest their main R&D resources in the "soul field", and the soul of car companies is "from hard to soft". Volkswagen CEO Diess, who takes automotive hardware to the extreme, publicly declared in 2019 that 90% of automotive innovation in the future lies in software.[2]
If the soul of car companies in the era of fuel vehicles is the manufacturing/tuning experience of engines, gearboxes and chassis, then the mastery of software algorithms for three electric systems, automatic driving, and intelligent cockpits is the soul of intelligent electric vehicle companies.
This is indeed the case: a car's ability to drive itself is largely determined by the chip as hardware and the algorithm as software. Chip iteration is slow, barriers are high, car companies have little willingness to develop themselves; software algorithm iteration is fast, consumers can directly perceive, and car companies generally focus on research and development in order to differentiate their competitiveness.
In order to defend the new soul, in the past five years, from new car-making forces such as Xiaopeng, Weilai and Ideal, to traditional car companies such as SAIC, and then to BMW and Audi with a longer history, an algorithm team of autonomous driving software has been formed. These teams, like Tesla, are opposed to packaging, except that their demand is not to develop everything themselves, but to decouple hard and soft.
For example, the original mainstream autopilot chip is a high degree of coordination between hardware and software algorithms, more like an exclusive, once the car company has no right to interfere in the independent kingdom; and the soft and hard decoupled chip, must be a hardware generalization, software support downstream secondary development programming platform.
However, mobileye's fame advantage is the "integration of software and hardware", its development ecology is known for being closed, and car companies want to replace the existing algorithm with self-developed automatic driving perception algorithms, on the one hand, it is not easy in technology, on the other hand, it will also be hindered by Mobileye.
In 2019, the Ideal One equipped with mobileye EyeQ4 chip was launched, and EyeQ4 became an "enclave" on the car, and the data collected by it was not shared with ideals. The helpless ideal is to install a data acquisition camera on the car. Weilai ES6, EC6 and other models that also use EyeQ4 chips, when developing NIO Pilot Pilot Assistance, because Mobileye's chip does not support car companies to modify the algorithm themselves, they have to write the self-developed algorithm into another chip, and the system operation is more complex and less efficient.
Three years ago, EyeQ4 was still a small sweetie that Weilai highlighted its product strength
"The world is bitter Mobileye for a long time". In addition to the traditional automotive semiconductor suppliers Renesas and Deyi, as well as NVIDIA designing GPUs, Qualcomm, which makes mobile phone SoCs, domestic startups Horizon and Black Sesame have also risen.
Although their experience in automatic driving chip algorithm accumulation and mass production is not as deep as Mobileye, they highly respect the needs of car companies for soft and hard decoupling - not only do not restrict car companies from developing automatic driving software algorithms, but also try their best to build tool chains to provide convenience for car companies to develop, transplant and update algorithms as much as possible.
NVIDIA relies on the AI developer ecosystem established by the CUDA development platform, which makes it difficult for car companies to apply artificial intelligence technology on its chips, so it has grabbed orders from Mobileye for new car autonomous driving computing platforms such as Weilai, SAIC, and Volvo. Horizon will be favored by the new Ideal One and more local car companies in 2021, relying on an open and flexible business model, as well as the local advantages of domestic startups to respond quickly and provide personal services.
This kind of "closeness" is often in the physical sense - the top brass of the ideal car self-driving research and development revealed a message, because the ideal has to quickly switch solutions, time is tight, and horizon engineers and ideal developers stay up all night. In contrast, Mobileye has only one sales team in the country.
And recently evolved into intellectual property: Not long ago, Horizon announced the opening of the IP of its self-driving chips to help car companies develop their own chips. He was so fierce that he even changed his own life.
Over the past few years, although Mobileye's chip shipments have grown on cumulative orders, they have been in the process of customer churn. When Mobileye realized that he had changed from a car company's sweet and sweet to a cow lady, mass-produced the EyeQ5 chip that supported soft and hard decoupling, the time had come to 2021.
And the lessons it has to make up are far more than soft and hard decoupling.
03/Hashrate power grab: High T (ops) means justice
In the process of Mobileye being betrayed by car companies, NVIDIA, which came from a graphics card, snatched the most customers from its hands, including but not limited to Weilai, Ideal, Weima, SAIC, etc. Nvidia can continue to grab orders, on the one hand, thanks to its perfect development ecology, and more importantly: in the era of automatic driving perception technology turning to AI, NVIDIA took the lead in demonstrating the hegemony of computing power.
Since 2016, the automotive autonomous driving perception algorithm has undergone an industry-wide change: the traditional machine vision algorithm that was popular in the past was based on manually set rules, more like an expert to carry out logical deduction; today's mainstream is a deep learning algorithm, using a multi-layer neural network to extract image features for recognition, relying on parallel computing of large data volumes, similar to decomposing a difficult problem into 100 simple problems, finding 100 primary school students to calculate, simple and crude but effective.
Mobileye's strength has always been the traditional machine vision algorithm, and has independently designed the chip architecture, taking the route of high-efficiency ratio, and does not attach great importance to heap computing power; and the most important thing about NVIDIA's original game graphics card is parallel computing power, which coincides with the needs of deep learning algorithms.
In 2018, the chip Paker computing power launched by Nvidia for autonomous driving is only 1Tops (tera operations, 100 billion operations per second), and In 2022, Orin, which will be installed on the NIO ET7, hashrate the hash rate soaring by 25,000% and exceeding 250T. By comparison, Mobileye's EyeQ5, which got on the bus last year, was less than 25T, with a top ten pieces.
The route chosen by the two is not 10 times the gap in practical applications, but the problem is that, like the running score that prevailed in the early years of smart phones, computing power has become the most intuitive data for consumers to measure the automatic driving ability of a smart car, and the perceptual understanding of "the greater the computing power, the more cattle", allowing car companies to enter the era of spelling power from the era of horsepower.
On the other hand, the industry does not have a unified understanding of how much computing power is needed for automatic driving, and the automatic driving function on intelligent electric vehicles must be continuously upgraded and updated, in order to avoid the embarrassing situation of "mass production is backward", car companies generally adopt the strategy of "hardware embedding, software keeping up".
For example, Weilai ET7 is directly in place, setting up 1000T+ computing power, requiring 4 Orin chips to work together on a huge board. If you switch to EyeQ5, all you may need is a cabinet that can fill the trunk, and the system becomes too complex, and reliability will be challenged.
In this future-oriented arms race, the higher the computing power of the chip, the greater the potential. Even if this is not the best choice in theory, the industry has entered a hashrate volume, computing power is justice, and high computing power can be bottomed out.
Mobileye once scoffed at this, and publicly compared The computational efficiency of EyeQ's chip with that of Nvidia, announcing that energy efficiency was beating Nvidia, but in the end, Mobileye had to bow its head under the hash rate race, and released The EyeQ Ultra in early 2022, with a hashrate of 176T, was still crushed.[3]
More importantly, this chip can only be mass-produced in 2-3 years, and Nvidia's Orin will be mass-produced this year, and Mobileye's high-computing chip for high-level automatic driving is lagging behind. Two months before that, Qualcomm had already used the high-computing power autonomous driving chip platform to snatch BMW, the biggest customer, from Mobileye.
04/Epilogue
After 2016, when the self-driving chip was planted to turn to the foreshadowing, the automotive industry experienced unprecedented turbulence, and the turn of technology and the rise and fall of the end market have witnessed the ups and downs of a number of supply chain giants. In the past two years, the lack of cores, lithium madness, and even war have made enterprises consider far more than the wind direction of technology and consumer preferences.
In a world where black swan events are frequent and supply chains are hit hard at some point, the criteria for car companies to judge suppliers have also changed, and a more open and transparent degree or lower risk of card neck has replaced cost and tacit understanding established by long-term cooperation.
Chaos is a ladder, someone takes the opportunity to climb up, and someone slips off the throne because of the situation.
Copyright Notice: In addition to publishing original articles, the manufacturing industry is also committed to the exchange and sharing of excellent articles. Reprint must indicate the source and author of the article; please reprint authorization please leave a message at the end of the article or in the background. All rights reserved, violators will be investigated.