Author| Utada
| Tiger Sniff Technology Group
The cover is from Visual China.
In the second half of 2021, the car factory's desire for lidar and high-computing chips and marketing gimmicks, as well as the strong attack of Nvidia's autopilot chip Orin, make it impossible for Mobileye, an Intel subsidiary with a market share of 70% of assisted driving chips, not to feel the strong pressure from the Chinese market.
And what is very interesting is that even though Mobileye released the L4 level self-driving chip this week, making a crowd exclaim that "this Israeli company is no longer 'possessed' by Intel, which is good at grinding, and is finally going to be stunned", there are still many people who are not optimistic about Mobileye's future. One engineer even made a very blunt point to us:
"There's nothing comparable, the car driver assistance chip giant is moving too late."
Mobileye CEO Shashua
Over the years, for many high-level autonomous driving (L3 and above) practitioners, NVIDIA is indeed "forever and the only god". Because 5 years later, the reality is still cruel and straightforward: stall hands, that is, there is no other similar product to choose from;
Since 2019, after the high-end autopilot trend has been forcibly pulled into the consumer-level automobile industry by Tesla and new forces, Nvidia has also become the best choice for old diehards.
After the release of the automatic driving chip Orin in 2019, the samples have successively reached the hands of autonomous driving companies and car companies in 2021, and many car companies have announced that future models will have Orin a place, which will inevitably crowd out mobileye's future market.
"If you look at it from a marketing point of view, unmanned driving is a gimmick, and the biggest gimmick is hash rate." I'm not sure if this computing power can really be fully utilized, but Nvidia offers an option that is not Tesla but can be compared to Tesla's hashrate, which Mobileye does not have.
If you look at this matter from a technical point of view, there is also a clear trend that car companies have begun to develop their own algorithms. If the algorithm is self-developed, then what everyone needs is an open hardware product, and Mobileye itself is not open enough, or it is not as complete as NVIDIA's open software toolchain. One software engineer believes that Mobileye has no advantages in terms of marketing and technical choices.
"In addition, from Intel's point of view, the CPU side of Intel and AMD are slightly inferior, how many resources are tilted on Mobileye, or how much choice space there is without touching Nvidia's patents, it does not seem particularly optimistic."
The above view is actually not new, but many people are deeply convinced.
Because we must recognize a cruel fact - that is, in the automotive industry, the gimmick of engaging in scientific and technological innovation has been gradually rationalized (the previous one is the mobile phone industry). Led by Musk calling the L2+ feature "fully autonomous," any exaggeration in the auto industry seems to be too much.
In addition, for NVIDIA, which has a super software ecology and tool chain, the Orin mass production in 2022 does not represent their actual level in 2025. And mobileye's EyeQ Ultra to be mass-produced in 2025 is compared to the Orin mass-produced in 2022, or the NVIDIA level in 2025?
If the two companies compare the hardware capabilities to carry L4 or above autonomous driving, it is very obvious that Nvidia wins.
However, adding a "consumer car" to the L4 is another matter.
Image courtesy of eeNews Automotive
Of course, it must be noted that the L4 autopilot chip has also been called by some analysts as "the last attempt for valuation before listing".
Because another big event related to Mobileye's business prospects in the next 5 years is the "Intel spin-off listing plan" that made the entire Wall Street focus on it last month.
The israeli small giant of assisted driving chips, which was acquired by Intel in 2017 for a superhero of $15.3 billion (about 42 times its sales in the same year), will independently IPO in 2022, valuing it at a whopping $50 billion.
Of course, Intel, the semiconductor giant whose market capitalization is now nearly stagnant, is a bit too obvious for the purpose of this split — to raise enough money to fill the pie of "investing at least $80 billion in fabs around the world" in 2021.
But the IPO plan also caused great controversy on Wall Street .
While Mobileye's annual revenue finally exceeded the $1 billion mark in 2021, many analysts have questioned the $50 billion valuation for revenue forecasts for the assisted driving business over the next 5 years, as well as the challenges facing NVIDIA and other emerging Chinese self-driving chip companies in the next 3 years. The more delayed the IPO, the less favorable it is to its valuation.
In addition, as mentioned above, how much intel has skewed resources for Mobileye, and whether Intel's traditional business capabilities will be more thoroughly exposed after splitting Mobileye (after all, AMD is eyeing it), are doubtful.
For example, as a subsidiary, Mobileye did not explicitly state that it used Intel's factory to make chips. According to public information, its early EyeQ series processors used the manufacturing process of STMicroelectronics (STM), while its latest generation of products chose TSMC.
Mobileye's road test car
Therefore, combining the brutal growth of China's auto market with global technology trends, although Mobileye's business revenue growth reached a historic 30% in 2021, and even in 2022, the rise is expected to be as high as 40%, but the praises of the Israeli company are endless, especially in China.
There are even articles that directly say that "Mobileye has fallen behind, and the era has ended"
As a result, the Israeli chip company, which has always been extremely arrogant and quiet in the image of the outside world, focused the biggest "firepower" of the speech at ces in 2022 on the energy consumption and power of the next generation of L4 level autonomous driving chips.
Of course, although this product is by no means an idea and achievement in the short term, the purpose and emotions of responding to market noise and competitor NVIDIA are about to come out .
Your strengths may be my weaknesses. And your weakness is exactly what I'm good at.
Not high in computing power? But don't underestimate me
Many times, we always put the most prominent things in a company to directly represent a company. For example, when it comes to Intel is the CPU (does anyone know that they are also very good at quantum chip research and development), Huawei is a mobile phone and communication device, NVIDIA is a graphics card, and Mobileye is a driver assistance...
Interestingly, however, if you look at Mobileye's annual CES presentations, their goal for 2025 has been to launch a "full self-driving service & product" (a term that no one who says I don't believe).
Image from CES 2022
The reason why Mobileye has always been considered conservative and "backward" is because they were founded in 2004 and finally became the Tier 1& 2, the premier assistant driver in 2014, and the identity is destined to reflect the will of the depot forever.
If anything is lagging behind, it's that standard automotive suppliers have always had to follow a car product update plan, not a technology update plan: each generation of new products is separated by several years, and it is a major upgrade to the previous product, and everything is based on "reliability"; not a CEO who insists that engineers modify a detail that can be iterated on (of course, this is not a good thing).
However, for the technology company, which has been established for more than 17 years despite its relatively new technology concept, they have made a significant change in the past two years - although they still sell assisted driving chips, they basically make all the key technical points that can be thought of around "L4-level driverless cars":
REM mapping technology (high-precision mapping), millimeter-wave radar and lidar systems, a standard that hopes to "unify" the attribution of safety responsibility for autonomous driving in the automotive industry - RSS responsibility sensitive safety model, and the unmanned taxi 2C-end service platform (MaaS) that we can't imagine in 2021.
It was almost the release of an L4 level autonomous driving chip.
The EyeQ Ultra, an autonomous driving chip product that just disclosed the basic parameters at CES this week, although built on 4 self-developed accelerators developed by Mobileye very early, quickly triggered the controversy of "insufficient computing power" - only 176Tops (processor computing power units, which usually require higher computing power for artificial intelligence training), which is 1/5 of the level of NVIDIA.
But unexpectedly, Professor Shashua, CEO of Mobileye, who has always been recognized as the israeli uncle with an extremely sharp mouth in the industry, showed his "self-knife" ability at the beginning:
Compared to the numbers of competitors (don't doubt, here NVIDIA), 176 sounds like a small number, probably only 1/5 of the power claimed by competitors. But the key is not only computing power, but more importantly efficiency. ”
He directly pointed out a point that a chip engineer has verified for years — only by deeply understanding and understanding what the interaction between software and hardware is, it is clear what kind of core is used to support what kind of algorithm.
The evidence comes from their experience in driverless taxis everywhere, because these cars are powered by 8 EyeQ 5 (their latest generation of mass-produced assisted driving chips), and 176TOPS is about the sum of 10 EyeQ 5 chips.
"L4 can be driven by a single-chip electronic architecture, and the system consumes less than 10W and costs hundreds of dollars. This is a major step forward in terms of cost and energy efficiency. These dimensions are the "seven inches" of the depot.
If you think about it this way, Mobileye, which will only have samples in 2023, is indeed behind NVIDIA, which has long occupied the initiative in the high-end automatic driving market; but the energy consumption, cost and efficacy of L4 consumer cars, as well as more than a decade of experience in car factory cooperation and linkage, can help Mobileye get some trotting opportunities.
On the other hand, although many media statistics NVIDIA Orin has received a large number of car brand orders, according to an experienced automotive chip engineer, although NVIDIA's products are indeed very strong, most car manufacturers still have various problems in the testing process:
On the one hand, the car factory simply does not have enough talent and ability to release all the chip performance, and on the other hand, Nvidia does not care too much about service capabilities.
"Just like in the martial arts novel, the Shangfang sword hanging there, you can get it, but you don't necessarily have the 'life' to use." There are very few engineers with software and hardware collaboration and tuning capabilities, let alone automakers. ”
What Shashua wants to emphasize is the current market feedback situation - in the case of low Tops, if the computing power is used very efficiently, it is not necessarily too high Tops products.
In other words, through good hardware and software design, a balance between efficiency and lean computing can be achieved. Is this a technical advantage? It is left to the automotive chip tuning engineers in the industry to judge.
This is actually easy to recall a large wave of mobile phone manufacturers in 2018, all taking "AI mobile phones" as a marketing gimmick, and directly linked to high-computing AI chips. But in fact, according to chip engineers at the time, in fact, there are far less AI functions on mobile phones that can use the latter's computing power.
Another overlooked strength – the radar system
Unfortunately, because Mobileye always presents a "proud and low-key temperament" to the outside world, the technical exposure classes they only open at CES every year are easily overshadowed by too many exaggerated and gimmicky concept technology products. Of course, the proud Israeli technocrats don't care either.
Last year, someone asked what Nvidia's next most representative and influential product after the CPU was, and some analysts actually said a particularly interesting and alternative answer - it may be Mobileye's lidar chip.
Lidar, the automotive industry's hope of leaping to the phenomenon-level parts of the L4 self-driving car, in addition to triggering wave after wave of entrepreneurial boom, is also an in-vehicle product firmly held by Huawei and DJI. However, not many people understand that Mobileye released a lidar chip last year, and it is a very difficult FMCW (Frequency Modulation Continuous Wave) technology route product.
Last January, we were shocked by the news that they had released the FMCW lidar chip. You know, a company that does perception systems has entered a hardware field that seems to be unrelated to the original market, and it has chosen a relatively difficult path, which is one of the few in the industry.
In addition to its original chip design and production experience, Mobileye is actually on the same line as lidar FMCW chip startups such as Aeva, Blackmore, which was acquired by Aurora, and Voyant Photonics, which have even just received Series A financing.
But from the series of actions of Mobileye this year, in fact, we have become more and more clear about the purpose of Mobileye.
In addition to maintaining close control over the market for assisted driving technology, the self-developed lidar, millimeter-wave radar, L4 autopilot chip, and self-driving taxi service are all based on the perspective of the depot, deep horizontal interleaving experiments "how to significantly reduce the cost of sensors, and maintain low power and long life, under the premise of achieving high-end automatic driving."
Similar to Shashua's emphasis on the low power consumption of the Mobileye L4 autopilot chip, the biggest advantage of FMCW technology lies in the cost and stability of chip-level products, as well as the optimal balance between human eye safety and ranging performance.
However, due to the extreme difficulty of implementing this route and the current results are few, there should be no more than 10 companies developing this technology in the world.
Of course, the domestic head lidar company may not care - after all, many new models in China from 2022 to 2024 have long been fixed to use whose solid-state lidar.
But in fact, the "best ratio of technology, vehicle regulations and cost" pursued by the most important in-vehicle market segment at present is far from final. Even the domestic lidar "Two Dragons and Two Little Dragons", which has ordered from automakers, is constantly developing different technical routes internally.
In any case, long life, stability and maximum security are things engraved in the bones of mainstream car companies around the world, and "chipping" will be the most reliable form. Therefore, the opponents that lidar startups will encounter will be far more powerful and terrifying than those on the market today.
However, the LiDAR silicon photonic chip that Mobileye said will be mass-produced in 2024 is actually only one of the branches of the company's internal radar development team.
The three new products to be released in the future — the L4 autopilot chip (left), the redefined imaging radar (center), and the FMCW lidar chip — are described by Shashua as game-changing products in the future.
Through the L4 level road test vehicles and map mapping vehicles in various countries, they not only collected the necessary software training data, but also did the performance innovation experiments of camera perception systems and radar series products.
Interestingly, they don't want to fuse cameras with sensors such as radar (which is the mainstream view held by the industry at present), but firmly believe that the real redundancy concept of L4 should be cameras, millimeter-wave radar, and lidar as completely independent subsystems.
"If you make such a redundant system consisting of two subsystems, you can not only improve robustness and mean time between failures, but also make it easier to verify the true efficiency of the system."
What is more interesting is that the brain circuit of Israeli technologists "dare to think and dare to do" has been vividly experienced in their other lidar branch route .
You're just called lidar, but with millimeter-wave radar, you can't do what you can do? Why can't we redefine it?
The basis of this idea is the cost thinking of the depot - the cost of traditional radar is about 1/5 to 1/10 of the cost of lidar.
In fact, this idea was proposed by domestic millimeter wave radar companies as early as 4 years ago (when mechanical radar was the mainstream, the price was as high as tens of thousands of hundreds of thousands) and was far from being accepted by the car factory, but there was no follow-up after that. But the Israeli uncle just showed what they've accomplished with deep learning:
Through the training of neural networks, the data quality of traditional millimeter-wave radar output has reached almost the same level as that of lidar data;
Mobileye trained the neural network to use the data is the traditional radar input and lidar output, when the radar and lidar are put together, with lidar as the neural network output, the radar is mapped to the lidar, the effect on both sides is basically the same
Even the output data of the millimeter-wave radar has the same effect as the camera image.
It's hard not to think of the unforgettable hardware and software collaboration capabilities of Google when it released the Pixel phone — indeed, in the case of a sufficiently powerful algorithm, it is completely possible to make up for some performance deficiencies in the hardware.
But, as I said at the beginning, sometimes gimmicks are also important – many new sensors such as lidar are now part of brand marketing in China.
But in any case, since they have chosen to release lidar and L4 autopilot chips in the past two years, smart Israeli companies should incorporate many factors into their business landscape in the next few years.
In short, without exaggeration, never underestimate the company that has fully mastered the system linkage between the software and hardware of automatic driving, and never underestimate the innovation of the Israelis.
Write at the end:
So, will Mobileye rely on Intel's powerful chip manufacturing capabilities in the future?
Although this part of the advantage has not been highlighted before, the IDM2.0 strategy released by Intel in May 2021 clearly regards Mobileye as a key part of its future strategic map:
Invested $10 billion in a chip factory in Israel and $400 million to expand mobile's autonomous driving technology research and development center.
In addition, in addition to 1 chip delivered by Mobileye for more than a decade, in fact, in the complex commercial market, more importantly, it has paid a huge cost of running-in for the car factory, and the entangled interest ties they have established with various agents around the world, especially in China.
"It is undeniable that NVIDIA is very strong, but what is revealed in the media is not in line with the actual situation that Mobileye has experienced in the industry on a global scale." A supply chain source pointed out that Nvidia and the current so-called domestic L4 autopilot chip company, although the momentum is fierce, has not shaken the mobileye market. The attitude of the depot is actually very subtle, because the depot can never stand behind a chip factory.
"But Mobileye does need more openness, whether it's software or disclosure. They typically belong to companies that have done a lot, but have long been used to being hidden champions, and hardly mentioned. ”
But once it comes out to speak publicly, it is bound to be targeted, and even cause an uproar.
It has to be said that Shashua, the Israeli uncle, his straight temper, extraordinary technical analysis and eloquence and satirical writing not only highlights Mobileye's pride as a subdivision industry giant, but also proves that the competition between the strong is the most likely to trigger industry changes.
Of course, it is also the most interesting.
Just like in early 2019, Shashua's "NVIDIA 'plagiarized' Mobileye Security Sensitive Liability Model (RSS) inside and outside the paper vibration industry" - "Innovation Requires Originality" (the following is an interception):
"It is said that 'imitation' is the most sincere form of flattery. Our innovations have not gone unnoticed, and many people agree with the same concepts we pioneered. In particular, there is an industry player who is used to following in our footsteps, and today I would like to clarify the facts about its latest imitations.
The RSS we presented was published in an academic paper in 2017. We have publicly shared all the technical details and mathematical knowledge behind RSS because we believe that the safety concept of autonomous vehicles should not be unique, and the industry should work with the government to study the meaning of safe driving.
We openly invited the entire industry to contribute their ideas on how to improve RSS. In 2018, Nvidia responded to this call to communicate on relevant cooperation matters, and we are very happy. But then we were confused when Nvidia withdrew from the partnership.
Imagine how surprised we were when NVIDIA CEO Jen-Hsun Huang announced an "unprecedented" self-driving car safety model (SFF) last week. Curious about NVIDIA's 'firsts', we eagerly read the SFF white paper, but the result was very high - we felt like we were looking in the mirror.
If imitation is the most sincere compliment, then NVIDIA must have rated us highly. It's clear that Nvidia's so-called 'first' safety concept is a replica of the RSS model we released two years ago. In our opinion, SFF is just an inferior version of RSS.
To some extent, if there is any innovation in this regard of NVIDIA, it is only an innovation in language. ”
Please continue to fight, so that it is conducive to the real progress of the industry.