Works of the New Eye Industry Group
The author | Ruan Xue
Edited | Sang Mingqiang
For a long time, automatic driving has been a regular visitor at the press conference of major new energy vehicle companies.
Different from the exciting rhetoric on the PPT, the initial freshness of automatic driving has been slightly auditory fatigue, coupled with the frequent occurrence of safety accidents in some car companies, consumers who understand the principles of automatic driving technology have begun to return to calm. The question that people are thinking about has also changed from whether automatic driving is safe at the beginning to questioning the authenticity of the proposition of automatic driving.
The reason for this is not that people can't wait for new technologies, but that the waiting time is too long.
"I knew from the beginning that Autopilot was a problem that was difficult to solve, but I didn't expect it to be so difficult to solve." At the beginning of 2022, when Musk made a guest podcast on Lex Fridman to talk about automatic driving, he also admitted that the problem had exceeded the difficulty of the original plan.
This year, it is the 8th year that Tesla has studied automatic driving, but if according to the strict classification of automatic driving, its technology is still stuck at the L3 level, and although FSD reveals Tesla's expectations for Full Self-Drive, it also reveals the helplessness of development.
However, Musk still seems to be optimistic, in his view, Tesla FSD will be able to achieve L4 level self-driving in 2022, and also predicts that the accident rate of FSD is at least 2-3 times lower than that of humans. Try to raise people's expectations and imagination of autonomous driving again.
But obviously, it's not enough to just give people imagination, and FSD shouldn't just give Tesla imagination. Compared with the symbolic fully autonomous driving, people living in reality have repeatedly underestimated the commercial value of FSD, as if it were a fate that had been predestined since its birth.
01
Schrödinger's self-research
"Importance and value potential", this is the word Musk emphasized FSD during the performance briefing call after Tesla released the fourth quarter of 2021 earnings report. After that, Tesla opened up the "fully autonomous driving" function (FSD) to other car companies on his Twitter, and he also declared that if other car companies want to develop similar functions through self-development, it will take at least 5 years.
How long did Tesla take? Before we talk about FSD, let's talk about Autopilot.
Also on Twitter, autopilot was built from a vision by Musk in 2013. In the eyes of industry insiders at the time, Autopilot was just an L2-level automatic driving system, strictly speaking, an assisted driving system, which meant that it could assist car owners to complete basic driving tasks under specific conditions, but it was by no means completely autonomous driving.
In the official function setting, it has not thought of letting it carry the mission of full automatic driving, which has the core disagreement between Tesla and the first partner Google and the suspended cooperative semi-autonomous driving project.
Turning to Mobileye became Tesla's helpless choice at that time, so that in 2014, Tesla AP HW1.0 came out, equipped with a forward-facing millimeter-wave radar, a forward-looking camera, 12 ultrasonic radars, a set of high-precision electronic auxiliary braking system (remote brakes) and a motherboard equipped with Mobileye Eye3 chip, although the initial function setting was completed, but it did not fully meet Musk's expectations.
Until the end of October 2015, Tesla's autopilot officially entered the historical stage, and subsequently, every technological update, the arrival of full autopilot will become Musk's new Flag. Also in the same year, Tesla officially released the V7.0 version of the in-vehicle system in China, adding features for autonomous driving, including automatic lane keeping, automatic lane change and automatic parking.
However, the cooperation between the two companies did not last long, on the one hand, dissatisfied with Mobileye's slow progress, on the other hand, dissatisfied with Mobileye's use of Tesla owner driving data to improve the chip algorithm. As a result, in July 2016, Tesla stopped cooperating with Mobileye and announced that all vehicles will use NVIDIA autopilot chips.
At the same time, Tesla's self-research project has also quietly started. Unlike NVIDIA's autopilot chip solution, which mainly relies on the GPU with greater power, Tesla has focused on the team's self-developed NPU (neural network accelerator), which is mainly used for image processing, accounting for the largest proportion of space in the integrated chip and undertaking the largest workload.
According to the rough statistics of New Eye, the FSD chip has gone through 18 months from design, to testing and mass production.
In April 2019, Tesla successfully launched its self-developed FSD master chip on the Autopilot HW 3.0 platform, realizing the vertical integration of autopilot chips + neural network algorithms. Looking back at tesla's autopilot upgrade road, the birth of FSD in the basic strategy of Autopilot hardware first and software update is to paint its intelligent color deeper.
The AP1 released in 2014-2016 has initially realized auxiliary driving functions such as automatic cruise, automatic steering, automatic auxiliary lane change, automatic parking, and summoning; the AP2 in 2018 upgraded the summoning function and added automatic assisted navigation driving; until 2019, the launch of FSD, Tesla further realized intelligent summoning and new traffic lights, parking sign recognition and control and other functions.
Regarding the establishment of FSD, in addition to the unsatisfactory speed of technology updates of cooperative manufacturers, there is also Tesla's pursuit of autopilot solutions, as Pete Bannon once said after the mass production of FSD, "The reason why Tesla began to develop FSD is that we found that there is no chip on the market designed for autonomous driving and deep neural networks from the bottom." ”
02
Self-driving puzzles
The ability to free your hands is perhaps the biggest misconception about autonomous driving right now.
In fact, no self-driving R&D company has been able to achieve full autonomous driving, and the driver's active monitoring is still in the first place. FSD is essentially still an evolutionary version of the automatic driver assistance system, which automatically assists steering, acceleration and braking in the lane under the condition of active driver monitoring, and the ultimate role is defined as reducing boring driving operations and improving driving pleasure.
This is obviously different from people's imagination of autonomous driving, but there is a logical inversion in the practical application and the realization of technological breakthroughs.
The core technology of autonomous driving is divided into three parts: perception, planning and control. Placed in the specific driving scene, perception is equivalent to people driving when looking at the road, obtaining road information, planning is equivalent to processing road information in the brain and output driving decisions, and control is equivalent to people manipulating hands and feet to drive. Compared with the operation of the operation, perception and planning are the biggest problems of automatic driving.
Regarding the solution of perception, in the field of intelligent cars divided into two major factions, one is lidar and other high-precision components for road condition recognition, improve the perception of the car, domestic manufacturers gradually began to use camera + high-precision map + lidar comprehensive solution; the other is the Tesla as the representative of the camera faction, firmly believe in the power of vision, that people can effectively distinguish the road conditions because they can rely on the eyes to capture information, while the cost is low, can be through the algorithm upgrade iteration to achieve various functions.
Compared with the high cost of perceptual elements and the high recognition error rate, Musk believes that "if humans rely on their own vision to identify their surroundings, then cameras can also achieve human eye function." "After the environmental data ingested by the camera is processed by the visual algorithm, the system will train itself through the deep learning model, so as to achieve a full range of cognitive road conditions and improve the accuracy of system control."
FSD inherits such a visual solution, relying on ingestion data to improve the performance of the ADAS system, it can be said that Tesla's massive driving data has become its nourishment. The advantages of Tesla's FSD system are also reflected in high-speed image processing, NPU, and SRAM.
At present, in the processing performance of high-speed images, Tesla's image processor SIP has reached the fastest consumer-grade video transmission DP1.4 standard; The calculation results used to store the NPU, the SRAM capacity with the caching function has reached 2TB/s, which theoretically meets the requirements of the L5 level of automatic driving.
In the era of automakers have a hardware power, Hardware 3.0 version integrated 2 FSD chips, finally lipped 144TOPS, will be launched in 2022 self-developed Hardware 4.0 version, is expected to perform will be 3 times the HW3.0, can reach 432TOPS, obviously Tesla is still in front of a group of friends.
Correspondingly, Tesla's business model is also at the forefront.
03
Software-defined cars
Cars are becoming more and more like walking computers, not only in the driving experience and technology upgrades, around the commercialization of autonomous driving, FSD is an unavoidable topic.
Tesla adopts a business model of "self-developed systems and chips + car manufacturing", which can not only reduce long-term costs, but also obtain benefits from vehicle sales. As a simple example, Tesla has already achieved a gross profit growth rate of 105% in 2021 alone.
As the self-driving capabilities continue to upgrade, so does the price of FSDs, as if citing Musk's predictions about the commercial potential of FSDs, and the tendency of software-defined cars to become more pronounced.
Since Tesla opened the AP system in 2015, the price is $2500 / set, and then raised to $5000 / set. By March 2019, users can pay an additional $3,000 for an FSD (which doesn't include any features at this time) in addition to the $5,000 EAP (Enhanced Autopilot) package.
In April of the same year, Tesla canceled EAP and moved the EAP function to the FSD, which increased the price to $6,000 per set, and users could get the Basic Autopilot (BAP) function for free. In October 2020, the FSD Beta version launched a fully autonomous driving test function equipped with urban roads, and the price was raised to $10,000 / set, and in January Tesla FSD rose again to $12,000.
Every time there is a renewal, the price will rise, which has become the norm for FSD.
Figure: Tesla FSD Fee Increase Table (Source: Tesla, Guosen Securities Economic Research Institute)
In July last year, Tesla launched the FSD subscription package to deepen the impression of car software, one-time payment and subscription payment of the dual track operation, to a certain extent, open up the new imagination of car manufacturers on commercialization.
For example, Volkswagen's marketing executive once conveyed another payment model message: Volkswagen can sell fully autonomous vehicles to customers at a price of 7 euros per hour (about 55 yuan) in the future, that is, pay-per-use mode. The variety of payment models for autonomous driving services has also become a way for car companies to find a balance between cost, profit and user consumption level.
Compared with the ceiling effect of hardware development, the upgrade of software and service systems provides a new profit model, and the profit space brought by it has jumped out of the business logic of traditional cars.
Take FSD's one-time pay model as an example: Revenue is expected to grow rapidly from $950 million in 2020 to $14.176 billion in 2025, with a CAGR of 72%. Based on the FSD net profit margin of 55%, the profit margin of FSD in 2025 has reached nearly 7.8 billion US dollars, and the technological update behind the price has also become the hard truth of profit growth.
Under this logic, it is not so much that people underestimate FSD, but rather that the expectations and misunderstandings of automatic driving are too high to make people ignore technological upgrades, and Musk's collapse of Flag again and again has also caused people to feel disappointed in automatic driving.
But it cannot be denied that in addition to FSD, Tesla's construction in the automotive ecosystem is also transforming the car from a "walking computer" to a "walking third space", and those that appear on PPT are gradually being implemented, but one thing needs to be recognized: no one can give a final deadline, even if that person is Musk.
But like Tesla running down the road, FSD is always on the road and never even has an end.