Within the past 24 hours, multiple heavyweight messages have been transmitted from the intelligent driving track.
Tesla began launching the FSD (Fully Autonomous Driving) V12.3 version in North America on March 18th local time. The biggest highlight of this upgrade is the introduction of Tesla CEO Musk's proud "end-to-end neural network" technology. This transformation is known as the "technology that changes the rules of the game.".
In addition, NVIDIA CEO Huang Renxun also announced at GTC on the same day that the centralized in car computing platform DRIVE Thor will be equipped with a new Blackwell architecture, and many Chinese car companies including BYD, Aion, Ideal, and Xiaopeng will be equipped with this AI platform.
Under the joint catalysis of multiple factors, A-share intelligent driving concept stocks have set off a limit up trend. On the morning of March 19th, Tianmai Technology and Wanji Technology successively hit the limit up by 20CM, while Sanyou Technology, Jinyi Technology, Power New Technology, Derun Electronics and others also hit the limit up, and the stock price of Desai Xiwei also rose significantly.
"Techniques for changing game rules"
On March 18th local time, Tesla owners in North America received updates and notifications for the FSD (Fully Autonomous Driving) V12.3 version.
According to user feedback, the driving experience of V12.3 has significantly improved compared to the previous V12.2.1 version, which has attracted widespread attention in the industry. It is reported that FSD V12.3 is the first official V12 version pushed by Tesla to all FSD users, and it is expected to achieve full coverage in the near future.
The biggest highlight of this upgrade is the introduction of Musk's proud "end-to-end neural network" technology. The application of this technology in the car means that Tesla's control logic has been transformed into neural networks for processing, which is known as a "game changing technology.". In Musk's words, this is a major update of the version, and can even be called the V13 version.
In fact, at the beginning of this year, Tesla once pushed an update, but it was limited to employees and beta testers, including some early FSD beta testers.
It is understood that starting from FSD V12, Tesla's autonomous driving has entered a new stage: there are no rule codes, only neural networks. Compared to previous models that typically incorporate human written "if else" rules, the entire end-to-end algorithm is almost entirely constructed using neural networks, with sensor perception information as the input and control results as the output.
Namely, various data are input by the vehicle end sensors, processed by AI algorithms themselves, and finally output driving decisions to control the vehicle. During this period, both training and practical operations were driven by data.
In the industry's view, the advantage of such algorithms lies in their high flexibility, good adaptability, high upper limit, and the ability to drive in a highly anthropomorphic state, greatly improving the driving experience, and even potentially surpassing human drivers in driving ability.
Tesla's comprehensive self-developed autonomous driving system
Caitong Securities pointed out that Tesla's road to autonomous driving is divided into three stages, and it has now entered a comprehensive self-developed stage, achieving comprehensive self-development from the underlying hardware to the upper software.
The organization believes that Tesla's technological innovations in the intelligent driving industry mainly include: shadow mode, which lays the foundation for Tesla's real data acquisition; HydraNet reconstructs the network structure of autonomous driving target detection to improve algorithm efficiency; BEV+Transformer, using BEV to upscale and collect 2D images to form the vehicle's own coordinate system; Occupy the network, supplemented with object height recognition and unlabeled obstacle recognition on the basis of BEV; End to end, based on deep neural networks, closer to real human driving.
Open source Securities also believes that building FSD V12 requires the collaboration of data, algorithms, and computing power. Firstly, on the data side, training algorithms is of utmost importance. How to collect massive amounts of effective data and train the model to exhibit corresponding driving behaviors in corresponding scenarios tests the developer's technical ability.
The computing power side is the foundation for training autonomous driving models. Tesla has a clear goal, relying on NVIDIA's GPU and its own Dojo supercomputer, to achieve 100EFlops of computing power by the end of 2024. Such a large-scale computing cluster provides a good soil for end-to-end model training and rapid iteration, which is essential for achieving end-to-end autonomous driving.
The institution also stated that on the algorithm side, end-to-end algorithms need to rely on the foundation of previous modular algorithms, and it is particularly important for developers to build good algorithm modules and systems. At the same time, algorithm clipping and training are also the focus of creating a perfect end-to-end algorithm.
NVIDIA's new GPU architecture "getting on board"
On the same day, during a GTC keynote speech, NVIDIA CEO Huang Renxun announced the new GPU architecture "Blackwell", and the GPU chip B200 based on this architecture adopts TSMC's 4NP manufacturing process. NVIDIA stated that it can achieve AI training and real-time LLM (Big Language Model) inference on a billion dollar parameter model.
In his speech, NVIDIA CEO Huang Renxun referred to Blackwell as the "engine driving a new round of industrial revolution.".
Huang Renxun also announced that NVIDIA's centralized in car computing platform DRIVE Thor will also be equipped with a new Blackwell architecture.
Several domestic electric vehicle manufacturers have showcased their next-generation AI models equipped with DRIVE Thor on GTC, including many Chinese car companies such as BYD, GAC Aion, Xiaopeng, Ideal Automobile, and Jike, as well as autonomous driving platform companies such as Wenyuan Zhixing.
As early as 2015, Nvidia entered the field of in car computing platforms and launched the first generation autonomous driving computing platform DRIVE PX and Tegra series in car chips. Later, Xavier chips and Orin chips were successively released. In 2022, as the growth rate of the automotive business slows down, Nvidia launched a new generation of autonomous driving computing chip DRIVE Thor. At that time, it announced a single chip with a computing power of up to 2000 TOPS, which is 14 times the computing power of Tesla's FSD chip.
According to Fraser Sullivan statistics, Nvidia's shipment volume accounted for 82.5% of the global market share of high computing power autonomous driving chips in 2022. According to Huang Renxun's plan, the future automotive business will become Nvidia's three pillar businesses alongside data centers and gaming.
Domestic autonomous driving acceleration landing
The two giants have made significant moves in the field of intelligent driving, and A-share secondary market concept stocks have also taken action. On the morning of March 19th, intelligent driving concept stocks rose successively after opening. As of the close of noon trading, the stock prices of several listed companies have risen sharply to the limit.
It is worth mentioning that autonomous driving is currently showing an accelerating trend in China. Wang Lei, Deputy Director of the Management Committee of Beijing Economic and Technological Development Zone, recently stated that currently, the high-level autonomous driving demonstration zone in Beijing has issued road testing licenses to 29 testing vehicle companies, and the mileage of autonomous driving testing has exceeded 25 million kilometers.
He revealed that by June this year, the demonstration zone would cover an area of 600 square kilometers, covering Tongzhou and Shunyi. At present, further planning for Phase 4.0 can cover most areas of the Pingyuan New City between the Fourth and Sixth Ring Roads in Beijing, and it has the basic conditions for preliminary promotion and comprehensive commercialization.
Yuan Fei, member of the Standing Committee of the Hefei Municipal Party Committee and Deputy Mayor, also stated that Hefei has built the first city level integrated supervision platform for vehicles, roads, clouds, networks, and charging in China, achieving centralized management of vehicles, smart transportation, and infrastructure, and connecting over 150000 new energy vehicles.
He stated that he will promote the development of industrial scale, commercialization, and integration, and strive to open more than 30 scenarios by 2024, with over 400 intelligent network joint testing vehicles on the road.
Gu Huinan, General Manager of GAC Aion, recently predicted at the China Electric Vehicle Hundred People Conference Forum (2024) that with policy support and the development of enterprise software and hardware, L3 level intelligent driving will explode in 2024, and L4 level intelligent driving is expected to be launched in 2026.
转播
