Nvidia and Horizon compete for car chips, and new forces follow Tesla's solo flight.

Text: Tan Qing said AI Author: Zheng Driving

"At the beginning, the propaganda was the chip in the car machine, and then changed the mouth to say the communication module, and now it continues to be changed to a remote module, what will it become next time?" I didn't even know that my car had so many more features. A good cat owner of Euler expressed his dissatisfaction with the manufacturer.

Recently, according to the financial network, the Great Wall Euler car good cat was suspected by the owner of the car to steal the chip for many days. The person in charge responded, "The intelligent service processing platform in the euler good cat parameter configuration table, the Qualcomm chip mentioned is a chip used in the remote control module. Some professionals said that Euler's response this time was tantamount to acknowledging the fact that Qualcomm cockpit chips were not used...

Tan Qing said that AI believes that in this seemingly ordinary car sales dispute, behind it actually reflects that the automotive industry is quite fierce in the arms race around intelligence, and is in the "division war" stage of chip research and development.

Intelligent, networked cars need to be in an interactive state at any time, requiring a lot of data processing and computing, conventional chips can not be applied to the future of the car, with a strong computing power of AI chips demand is increasing day by day. After all, car companies that can develop their own chips are a minority, and more car companies need to weigh cost, power consumption, computing power and other factors when cooperating with suppliers such as software and chips.

However, the more realistic problem is that the speed at which car companies launch new models is inconsistent with the development speed of chip manufacturers, who should accommodate whom? In this chip-centric automotive intelligent transformation, car companies and AI chip manufacturers are looking for a balance point that will benefit each other as much as possible, but such a process will inevitably have some variables.

Nvidia brute force overwhelms Mobileye, and Horizon differentiated architecture quickly catches up

After the advent of the era of intelligent cars, automotive electronic architectures have become more and more complex, and the overall evolution trend is from multiple discrete systems to more centralized domain control systems, and then to the future unified central computing platform. In this process, autonomous driving AI chips have become the core technology and have attracted many companies to enter.

These companies mainly have consumer electronics chip giants that do not cross borders, such as Intel, Nvidia, Qualcomm, etc., as well as more vertical automotive AI chip companies such as Mobileye, Horizon, and Black Sesame, as well as "self-developed" car companies represented by Tesla. Among the most representative of them, Mobileye, Nvidia and Horizon, have very different ways of playing, as follows:

* "Algorithmic" Mobileye

Mobileye started early in the field of intelligent driving, in the L1 ~ L2 low-level assisted driving stage, Mobileye occupies a dominant position, in 2017, after the acquisition of Xilinx Mobileye became a popular industry hegemon, its auxiliary driving AI chip market share reached 70%.

However, Mobileye seems to have prematurely begun the expansion of vertical integration, transforming the manufacturing model from Fabless to IDM, that is, not only participating in the design of chips, but also adding tapeouts, packaging and other services to achieve the integration of soft and hard chips.

However, the concept set by Mobileye does not seem to be bought by car companies with self-developed dreams, and the most typical representative is Tesla. Tesla purchased Mobileye EyeQ3 in the early days, and then the mobileye-led development rhythm could not match Tesla's needs, Tesla turned to the high-computing NVIDIA chip platform, and the current Tesla self-developed FSD has been mass-produced on the car, completely breaking free of the shackles of Mobileye.

Mobileye's former leading edge in ADAS has met the needs of car companies for rapid mass production, and also saved research and development funds for car companies. However, at present, ADAS technology is no longer new, different car companies have the demand for differentiated intelligent car products, and Mobileye's dilemma is how to deal with the conflict between meeting the personalized needs of different car companies and scale effects.

* "Hash power" Nvidia

NVIDIA is a leading manufacturer in the consumer electronics chip industry, maintaining an overwhelming share in the independent GPU market, and when other automotive chip companies try to find ways to use algorithms to make up for computing power, NVIDIA holds high the banner of "software and hardware decoupling" and becomes a strong force hostile to Mobileye.

Suddenly, this kind of "vigorous miracle" barbaric play has allowed car companies that are not satisfied with Mobileye's homogeneous autonomous driving AI chips to find better alternatives, coupled with the "buried" of redundant computing power means support for OTAs, Nvidia has indeed caused a lot of pressure on Mobileye. Before 2019, NVIDIA rode the dust on the high-computing automatic driving chip.

Compared with Mobileye, NVIDIA's advantages are not only higher computing power, but also the "openness" brought about by "software and hardware decoupling", and this "openness" is also a double-edged sword.

For example, tesla switched from Mobileye's EyeQ 3 to NVIDIA's Drive PX 2 at the end of 2016, and once had the problem of downgrading AEB functions, because the perception algorithm of Drive PX 2 was not mature enough, and Tesla did not have time to adapt the algorithm.

At this stage, most car companies are difficult to have the same self-research capabilities as Tesla, and can dilute the sales of research and development costs, so deep ploughing algorithms have become Nvidia's "compulsory courses", and also determine how much NVIDIA can eat in the future chip market.

* "Differentiated Enabler" Horizon

Horizon positions itself as a Tier 2 supplier whose main products include both self-developed algorithms and chips that control computing power, and is a company with strong financing capabilities. According to the Tianyancha APP query, in the past six months, its C round of financing has reached as many as 7 times.

In 2017, Horizon released the Journey 1 chip, which is aimed at intelligent driving of cars; in July this year, It released Journey 5. Judging from the speed and performance of releasing chip products, Horizon is in the first echelon in China.

The AI computing unit of Journey 5 adopts the BPU architecture (third-generation Bayesian architecture), the main role of the BPU is to support deep neural networks, the versatility is not as good as the CPU and GPU, but it is more efficient than the CPU in software implementation. BPU, on the other hand, once produced, cannot be reprogrammed and must be used under CPU control.

Liu Xin (pseudonym), a perception algorithm engineer who works at Great Wall Motors, told Tan Qing that for cars equipped with Journey 5, the realization of its OTA function must use multiple chips, and the upper limit of integration is not as good as ASIC (dedicated chip), because the degree of integration determines the performance ceiling of the chip, and the cost performance of BPU mass production is not as good as that of ASIC chips.

From the chip architecture of Journey 5, it can be seen that Horizon's current product positioning is between Mobileye's EyeQ and NVIDIA's Orin, taking into account both computing power and algorithm, and the advantage is that it can be replaced faster, but it is difficult to reach the level of ASIC like Tesla FSD. Therefore, the BPU architecture has actually decided that the Journey 5 series may be more suitable for "second-tier" OEMs that need rapid iteration.

At the 13th China Automotive Blue Book Forum held on June 11, Horizon founder Yu Kai mentioned in his speech, "We believe that autonomous driving must be human-centered rather than machine-centric, and our goal is not to make machines more powerful, but to make people owners of cars." The real value creation here is that human-computer interaction and automatic driving should be inseparable and integrated. ”

In Tan Qing's view of AI, automatic driving is related to safety issues, the grading standards are extremely strict, and it is not suitable or allowed to make differentiation, while human-computer interaction is more of a category of intelligent cockpits, which is a user-driven demand, so it is suitable for differentiation. The binding of automatic driving and human-computer interaction for Horizon means that the journey of the BPU architecture 5 can enable car companies to achieve the differentiation of intelligent driving interaction experience, which is a precondition for Horizon to be based on Tier 2 for a long time in the future.

However, Tier 2's position in the intelligent car supply chain depends on a considerable number of car companies that do not develop their own chips, and there is a long-term demand for chip suppliers. In fact, the motivation of car companies to develop their own chips has actually been exposed.

New force car companies "darkness Chen Cang", where can the horizons fit?

Automotive chips carry functions such as automatic driving and intelligent cockpit, which are related to the advantages and disadvantages of intelligent and networked experiences, and for users, the integration of many intelligent functions is essential.

Just like the Windows system itself comes with native system optimization functions such as defragmentation and firewall, but because the functions are too numerous and the operation is too complex, most users are difficult to master, so these functions are integrated by software such as 360 Security Guard into garbage cleaning, system optimization and other functions.

The degree of cumbersomeness of the car and machine system is no less than that of the PC, so many functions involving intelligent experience also need to be integrated to reduce the learning cost of users.

Under the general trend of intelligent function integration, in order to be an excellent chip supplier, whether it is Tier 1 or Tier 2, in the fierce horizontal competition, it is bound to find ways to expand the right to speak and expand the business in the direction of "full-stack solutions".

Therefore, most chip suppliers are not willing to stay in the chip itself, but gradually extend to the software layer to build their own industrial ecology. As a result, chip manufacturers will form an unavoidable "cut-off" of the soul of car companies, and car companies will have to join the team of self-developed chips in the face of "predatory" chip manufacturers.

Interestingly, different car companies have completely different attitudes on the layout of self-developed chips.

The first "Ming brand" core is Tesla. Tesla was the first car company for NVIDIA's PX 2 chip, but in 2017, Tesla announced that it had begun to develop plans for self-driving chips. Since April 2019, Tesla's new cars have used self-developed FSD chips, replacing the NVIDIA solution that was originally relied on, and setting an example for many car companies.

In 2020, a number of traditional car companies have disclosed plans to develop self-developed autonomous driving chips. For example, in May, BAIC Production and Investment established a joint venture with Imagination Group (a British chip design company) Beijing Core Technology, and in October, Geely's Yijiatong and ARM China established Core Optimus Technology, both of which include self-developed autonomous driving chips.

Compared with the traditional car companies that generously announced the establishment of a joint venture company model, the attitude of the new domestic forces on whether to announce the plan to develop their own chips is somewhat subtle.

The most typical is the "Wei Xiaoli" three new forces, although the plan of self-developed chips has not been directly disclosed, but various "omens" indicate that the self-developed chip matter may be in preparation as a "spare tire plan", waiting for the time to be ripe to fly alone like Tesla.

For example, Li Bin once said at the Chengdu Auto Show: "Weilai will definitely establish a full-stack automatic driving capability in the future, and chip capabilities are a key part of it." According to the news reported by 36Kr, Weilai Automobile has set up a hardware team code-named "Smart HW (Hardware)" to tackle chip technical problems.

In addition to Weilai, Xiaopeng Automobile CEO He Xiaopeng said, "In 2021, we will increase investment in research and development, including hardware closely related to automatic driving. Ideal is also more cautious, founder Li wants to say that "the premise of doing a good chip is that the algorithm is highly mature", but in April this year, Ideal has announced plans to make the automatic driving function standard for new cars, which may indicate that the ideal algorithm has matured and is not far from making chips.

Tan Qing said that AI believes that in the process of software-defined car transformation, the consumer electronics industry chip giants such as NVIDIA and Qualcomm do have inherent advantages in entering the automotive chip, and it is easier to form the illusion of "monopolizing" the market in the short term. But the intelligent competition of car companies is a marathon, and more incremental markets also indicate more variables.

From a longer time dimension, whether it is NVIDIA, which enjoys computing power hegemony or Huawei and Qualcomm, which do not build cars, they will face challenges from car companies' self-developed chips.

Ps: Tan Qing said AI, turn left new energy vehicles, turn right smart driving, there is depth, there is temperature, the author Zheng driving, reprint please retain copyright information.