Five pictures to understand the autonomous driving AI semiconductor ecosystem

Recently, taking advantage of the fact that Shanghai is under quarantine and can only stay at home for five days during the holiday, I am thinking about understanding the technology and supply chain behind smart driving. In fact, it is mainly divided into two parts:

• 
  

• The core competition in hardware is semiconductors.

• 
  

• The core of the software is the AI ​​algorithm.

I happened to see some articles and pictures from Yole, which are really good. They can help us quickly understand the key semiconductor ecosystem of the core hardware of autonomous driving. So I have processed these pictures again here, summarized five pictures for relevant interpretation, and strung them together into an article on the semiconductor ecosystem of autonomous driving:

1. Major players in smart driving chips

2. AI visual processing technology chain

3. Examples of AI Vision Car Chips

4. Development of major players in automotive vision chips

5. How do automobile OEMs respond to the strategy of AI chips?

I hope to give everyone a general understanding of the semiconductor ecosystem for autonomous driving. The five pictures can be clicked to enlarge.

1. Major players in smart driving chips

The perception hardware assembly is familiar to our articles or industry insiders. Cameras, LiDAR, and millimeter-wave radar are the three most valuable core components that run through the entire system. The semiconductors behind them are:

• MCU - Micro Control Unit A microcontroller, or single-chip microcomputer, is a small processor without a system. The others below are all with a system called SoC. This is widely used in radars. Of course, automotive actuators all use ECUs, which are the initial instructions to process simple information. This technology content and manufacturing process requirements are relatively low, so there are many opportunities for domestic production in the current trend, and its application time will also be long. Currently, MCUs are basically dominated by Infineon and NXP.

• 
  

• FPGA - This is where Microsoft dominates. Its advantage is that it can start system software (SW) and hardware (HW) development at the same time, mainly Xilinx chips.

• 
  

• Vision Processor - This is actually an AI vision processor, which is mainly used in the current GPU market. For example, my previous article on the six mainstream automotive chips and solutions for intelligent autonomous driving mainly talked about these, as well as the entry of Ambarella, Texas Instruments, China's Horizon, Huawei, and Black Sesame.

• 
  

• Central platform and accelerator - This is actually similar to the previous one. The difference is that it is deployed more in the cloud rather than on the vehicle side to process massive data and learn, such as Tesla's Dojo.

Of course, the core technologies involved in SoC are chip architecture and manufacturing process. In terms of chip IP architecture, there is a very powerful company called Arm that can provide technology. In addition, there are very powerful manufacturing companies such as TMSC TSMC. China Semiconductor Manufacturing Corporation, the domestic supply unit for the Shanghai epidemic, can also provide manufacturing process, so chip manufacturing can also be done in China. In general, AI is based on vision and language. Vision can be complex and even partially applicable to speech, so next we will focus on visual AI.

2. AI visual processing technology chain

This diagram expresses the three steps of the visual technology chain, from data collection, data preprocessing, to AI computing. Writing here reminds me of an article that was widely circulated in the WeChat circle before, Tesla killed the ISP (Image Signal Processor). In fact, it is meaningless to discuss this issue. Is it advertising for Tesla? First of all, the preprocessing of visual raw data is unavoidable. If it is not processed separately by the chip, it will be put into the integrated chip processing. It is determined by the camera sensor technology. My previous article "How many cameras does a smart car use? What do they do? What is the principle?" roughly describes it.

ISP can be processed by a separate chip, a sensor, or a central processor by a program. The current chip supply chain also supports multiple processing methods. However, this needs to be viewed separately from the perspective of cost and performance. For example, the ISP of the four surround-view cameras on the vehicle can be concentrated on one processor for processing. It is relatively simple. Reducing the ISP of each camera can reduce costs. This logic was proposed in "Seeing the Secret of Tesla's High Profits through Self-Driving Sensors and Chips". The design of automotive electronic components can use the concept of automotive platform architecture design to simplify electronic components and modularize them in batches.

3. Examples of AI Vision Car Chips

Continuing from the above information, the AI ​​car vision chip example, under the current supply chain, you can choose a single ISP chip for about $3-5, or you can choose ISP and processing fusion to support simple things such as vehicle line detection. The last one is ISP processing and fusion, such as Ambarella's chip.

I don’t see any Chinese products in this range, and it seems that the relevant technology for black sesame has not been confirmed.

4. Development of major players in automotive vision chips

How will the automotive vision chip develop? Yole believes that ISPs will also make computing chips, and companies that make algorithms or chip IPs will also make computing chips. It's interesting that AI is the future, and everyone sees this future and wants to get on the fast track. And at present, Ambarella, Texas Instruments, Renesas, etc. all have AI vision processing chips.

5. How do automobile OEMs respond to the strategy of AI chips?

As the largest application scenario of AI at present, automobiles. What strategies do automobile OEMs have? Yole gives three answers:

• 
  

• Both software and hardware are developed in-house, such as Tesla.

• 
  

• Purchasing hardware and developing software in-house is the strategy of current mainstream OEMs, such as Volkswagen, Toyota, and Ford General Motors.

• 
  

• Purchasing both software and hardware, and purchasing both AI software and algorithms from companies such as Stellantis, many people think this is a bad idea. But think about my previous article "Volkswagen's New Auto strategy is upgraded again - becoming a vertically integrated mobile travel company". Volkswagen's new competitor is not ABB and Toyota but Stellantis. In fact, it is mainly because I think it stands at a higher angle and regards it as a functional commodity rather than everything. You can click "Stellantis (FCA and PSA)'s software-defined car strategy" to learn more.

Obviously, we cannot simply evaluate the pros and cons of each company. Only time can give the answer.

Summarize

Today I happened to see a recent interview with Google CEO. He believes that artificial intelligence (AI) is the future of all industries. His requirement for Google is that all work should give priority to AI. Automobiles are currently the most widely used place for AI. The two most basic aspects of AI are chips, algorithms and related tools. For China, AI is another battlefield after the Internet. Have we won the Internet? At least I don’t agree. Our Internet prosperity is more of a demographic dividend. The underlying chips and system operating environment tools all come from other sources. Of course, for automotive AI, I am happy to see that Huawei, Horizon Robotics, Black Sesame and other companies have the bottom layer of our chips and algorithms. We have a good start.

So I hope this article can give you a broad introduction to the autonomous driving AI semiconductor ecosystem. If my article can help you understand a little bit of the technology chain, I think it will be meaningful.