Autonomous driving chip industry trends

Comparison of common autonomous driving chips

At present, the autonomous driving market has clearly entered a period of stagnation, and there has been no actual action in the L3/L4 market. There are two main reasons:

First, autonomous driving is highly dependent on deep learning neural networks that do not have interpretability. The lack of interpretability means that it cannot truly iteratively upgrade. Waymo, recognized as the dominant player in autonomous driving technology, has been developing autonomous driving for 14 years, but has not made significant progress in the past 10 years. This is why.

Second, the potential of lidar has not been fully realized. The current application algorithm of multi-line lidar is still based on deep learning neural network. Human beings have never found the best way to use point cloud deep learning, so that lidar is almost indistinguishable from cameras. This is also The reason why Tesla does not use lidar is that the market value of most listed lidar companies has dropped by more than 80%.

At present, many L2 level autonomous driving problems cannot be solved by systems that rely on artificial intelligence deep learning. For example, when encountering unrecognizable targets or sudden stationary targets, stereoscopic binoculars can solve these problems, but they are not mainstream technologies.

On the other hand, the R&D costs of autonomous driving have not decreased due to the stagnant market. The R&D costs of autonomous driving are rising rapidly. If you want to retain programmers, you have to keep raising salaries. Furthermore, the threshold for self-driving chips is not high and is very different from mobile phone or computer chips. Major companies such as AMD, MediaTek, Samsung, and Apple can launch self-driving chips in minutes.

This is the embarrassing situation of the autonomous driving chip industry. The market launch is far away, but R&D costs are increasing rapidly. There are also a number of companies that are eyeing it. The self-driving chip industry has to change. The current trend is to combine self-driving and cockpit into one, or self-driving and self-parking. In short, it is difficult for pure self-driving to find a place. Another factor driving this trend is The driving force is changes in automotive E/E architecture.

It is recognized that the future automotive E/E architecture will be a central computing architecture, or Zonal architecture. That is, all applications are completed by the central computing platform, which is like a server.

When Nvidia released Thor, the successor to Orin, in September 2022, it announced that Geely's Jikrypton would use Thor to build a central computing platform to complete all functions including smart driving, automatic parking, and digital cockpits.

NVIDIA Thor central computing platform

Qualcomm has the same concept as Nvidia. It is said that according to Qualcomm’s plan, the first generation Qualcomm Snapdragon

There are no upgrade plans after Ride, namely SA8540+SA9000, until 2025. Qualcomm’s main future plan is to launch a platform suitable for L2.5 with lower prices, including the SA8530P that has appeared recently and the lower-priced SA7550P and SA7530P that will be launched in 2024. . Qualcomm’s highlight in 2023 is the SA8795P, Qualcomm’s first central computing platform that combines smart driving and cockpit. In 2024, there will be a second central computing platform that combines smart driving and cockpit, the slightly lower-grade SA8755P.

In addition to Qualcomm and Nvidia, there is also Samsung. Samsung's AUTO V9 has successfully entered the Volkswagen supply chain. Volkswagen/Audi/Porsche's cockpit system MIB3 is all Samsung chips. Samsung originally had an A-series product line for smart driving, but it seems that fewer customers have adopted it. Samsung currently promotes Exynos Auto.

V920 is also Samsung's first automotive 4nm process chip, which should include smart driving and cockpit functions.

There is almost no public information about Exynos Auto V920, but using mobile phone platforms to share R&D costs is a magic weapon for Samsung and Qualcomm. Samsung’s current 4-nanometer process chips are mainly Exynos 2200, and V920 should be a modified version of Exynos 2200. Considering that the vehicle can be air-cooled or even water-cooled, the CPU configuration of the V920 may be 4 Cortex-X2 plus 4 Cortex-A710, and the GPU is AMD’s RDNA2 architecture Xclipse, with an estimated computing power of 2.1TFLOPs. The AI ​​computing power is said to be twice that of Exynos 2100, which is 52TOPS, which is not low.

Nvidia, Samsung and Qualcomm are all very familiar with the cockpit field. For those who are not familiar with the cockpit field, they need to consider covering functions such as automatic parking, electronic reversing mirrors and DMS. This is Ambarella’s CV3.

Ambarella CV3 internal frame diagram

The main focus of Ambarella CV3 is intelligent driving, and it is also obsessed with stereoscopic binoculars, including hard-core stereoscopic binocular parallax and dense optical flow. Bosch and Continental are also fans of stereoscopic binoculars. Perhaps this is one of the reasons why these two companies chose Ambarella CV3.

Target applications of Ambarella CV3

Typical cockpit domain controller architecture

For the cockpit domain controller chip, the main consideration is to consider more video outputs and more powerful audio functions. With a powerful GPU, more video outputs are a breeze, and the SA8795 can support up to 16 displays. However, the standard version of Orin only has one DP video output, which is through Type-C. Obviously, cockpit applications are not considered, but the Orin version for cockpits should have more video outputs. Of course, more higher-speed Ethernet interfaces are also necessary.

Audio is the core function of the smart cockpit, covering in-car audio, voice recognition, e-Call, noise cancellation and echo cancellation, active in-car noise reduction, active road noise reduction, DSP sound effects and other applications. With the development of intelligent networking in automobiles, the requirements for audio development are becoming higher and higher, and more advanced and feature-rich infotainment functions need to be added to satisfy consumers' user experience. However, the traditional analog parallel audio signal transmission method is difficult to achieve a balance between increased functionality and lightweight vehicle (reduced weight and cost of cables).

ADI (Analog Devices Inc.) has launched the A2B (AutomotiveAudio Bus) car audio bus through the optimization of the audio bus, which can provide better audio quality than the traditional analog audio bus, while also greatly saving the weight and cost of the car audio wiring harness (approximately less 75%).

If you want to achieve good car cockpit audio, you must use the A2B audio bus to separate the audio part.

Typical audio applications in cockpit systems

A2B is a one-master, multiple-slave link. There can only be one Master, which is configured in advance during the architecture design stage. It is usually the Head Unit host. If the host hangs up and there is a T-Box in the car, to ensure safety, the T-Box can switch from slave to master to send data downward to ensure that the subsequent network works normally. If a slave in the middle of the link hangs up, the downstream link of the node will hang up, the link will return from the node, and the upstream function will remain normal. The A2B bus is purely digital transmission. The digital interface can eliminate the need for peripheral DAC/ADC conversion and has strong anti-interference ability, ensuring audio quality to the greatest extent and improving the user's listening experience. Under the A2B architecture, after the nodes such as car computers and MICs are finalized, additional nodes, such as ANC microphones, etc., can be expanded and accessed when the nodes have not reached their maximum capacity.

The most competitive in the future will be Qualcomm, Samsung and Nvidia, especially Qualcomm and Samsung. Both have a huge mobile phone market to share costs, and they also have rich experience in the cockpit field. Samsung Harman is the world's number one smartphone. Cockpit manufacturer. Nvidia’s cockpit customers are being taken away by Qualcomm. Mobileye lacks experience in the cockpit field, as well as experience in automatic parking and electronic reversing mirrors. Renesas's main customers are Japanese manufacturers. Japanese manufacturers have always lagged behind in electronic architecture, so Renesas pays more attention to costs than computing power. Japanese manufacturers have always worked together, and Renesas can steadily obtain orders from Japanese manufacturers.