At present, there are three types of companies that drive passenger cars automatically. The first type is similar to Apple’s closed-loop system. The key components such as chips and algorithms are all made by ourselves. Tesla does this, and some new power car companies hope to gradually embark on this road. The second category is open systems similar to Android. Some manufacturers do intelligent platforms and some manufacturers do cars. For example, Huawei and Baidu < / P > < p > have intention in this respect. The third category is robotaxi, such as waymo. < / P > < p > this paper will mainly analyze the feasibility of these three routes from the perspective of technological and commercial development, and explore the future of some new forces of car manufacturing or automatic driving enterprises. Don’t underestimate technology. For automatic driving, technology is life, and the key technology path is strategic path. Therefore, this paper also discusses the different paths of automatic driving strategy. < / P > < p > in the field of intelligent vehicles, especially in the field of automatic driving, Apple’s closed-loop mode can not only make it easier for manufacturers to optimize their performance, but also can more quickly give feedback to consumer demand. With the gradual failure of Moore’s law, it is not feasible to simply increase the number of transistors per unit area to increase the performance. And because of the limitation of area and energy consumption, the scale of chip is also limited. Of course, at present, Tesla’s FSD hw3.0 is only in the 14nm process, and there is still room for improvement. < p > < p > at present, most of the digital chips are based on the von Neumann architecture of memory and calculator separation, which makes the whole system system of computer. From software to operating system to chip, they are deeply affected. However, von Neumann architecture is not suitable for the deep learning that autopilot relies on, which needs to be improved or even broken through. < / P > < p > for example, there is a “memory wall” where calculators run faster than memory, which can lead to performance problems. Brain like chip design does have a breakthrough in architecture, but it may not be able to be applied quickly if it is too far across. Moreover, the convolution network of image can be transformed into matrix operation, which may not be suitable for brain like chip. < / P > < p > therefore, as Moore’s law and von Neumann architecture encounter bottlenecks, the future performance improvement mainly needs to be realized through domain specific architecture. DSA, proposed by Turing prize winners John Hennessy and David Patterson, is an innovation that leaps forward a few steps and then leaps forward. It is an idea that can be put into practice immediately. < / P > < p > we can understand the idea of DSA from a macro perspective. Generally, there are billions to tens of billions of transistors in the current high-end chips. How to allocate the functions, how to connect and how to combine these transistors has a great impact on the performance of a specific application. In the future, we need to build a “fast system” from the overall perspective of software and hardware, and win by optimizing and adjusting the structure. < / P > < p > take another example to talk about DSA. In fact, mobile phone and terminal ecology, to some extent, are also using the idea of DSA. For example, there are also GPUs on mobile phones, which process visual data separately. There are also neural network accelerators on mobile phones, which are for deep learning. Apple’s latest M1 chip on MAC also follows this line of thought. There are GPUs and deep learning accelerators. According to the evaluation, the performance of many specific applications has been greatly improved. < / P > < p > of course, the DSA road of smart cars will go more thoroughly and deeply, because automatic driving is a specific chip running a specific application, without considering ecological problems. < / P > < p > Tesla’s FSD HW is specially designed for autopilot, which can be optimized at both ends of hardware and software. For example, the proportion of convolution operation is very high, then the convolution operation can be parallelized and specially optimized, which can greatly improve the overall performance. At present, only one frame per second can be processed by this neural network, which can not meet the demand of processing a frame of image per second for the image of < 350 million frames per second. < / P > < p > to explain, according to this data, the total computing power of FSD HW 3.0 is 35g * 2100 = 73.5tops. Considering some rounding or accuracy errors of the data, it basically conforms to the 72tops data claimed by Tesla. < / P > < p > both cloud training and terminal reasoning can be well optimized by DSA. For example, Tesla’s dojo is aimed at cloud training, which not only involves algorithms to chips, but also needs distributed machine learning. FSD HW 3.0 is responsible for terminal reasoning, and its performance is stronger than that of competitors. < / P > < p > closed loop mode is conducive to rapid response to consumer demand. For example, the separation of algorithm and data will make it more difficult to improve. Some corner cases may need to adjust algorithms and data at the same time, so the coordination of different companies will become a problem. < / P > < p > in general, the closed-loop mode adopted by Tesla can not only help improve performance, but also respond quickly to consumer demand, which is the best solution for automatic driving at present. < / P > < p > many people believe that in the era of autonomous driving, there will be apple and Android in the field of smart phones, and there will also be heavy core software providers like Google. My answer is very simple. The Android route will not work on automatic driving because it is not in line with the direction of future smart car technology development. < / P > < p > of course, I will not say that Tesla Weilai and other companies have to make their own screws, and many parts still need to buy parts manufacturers. But the most important part that affects the user experience must be done by ourselves, such as all aspects of automatic driving. < / P > < p > in the first section, it has been said that Apple’s closed-loop route is the best solution. In fact, it also demonstrates that Android open route is not the best solution in the field of automatic driving. < / P > < p > smart phones and smart cars have different architectures. The focus of smart phones is ecology. Ecology means providing a variety of applications on the basis of arm and IOS or Android operating system. Therefore, Android smart phones can be understood as a combination of common standard parts. The standard of the chip is arm, Android operating system is on the chip, and various apps are on the Internet. Because of its standardization, both chip, Android system and app can easily become a business independently. < p > < p > the focus of intelligent vehicle is algorithm and data and hardware supporting algorithm. The algorithm, whether training in the cloud or reasoning in the terminal, needs high performance. The hardware of intelligent vehicle needs a lot of performance optimization for specific applications and algorithms. Therefore, only doing algorithms or chips or operating systems will face the dilemma of performance optimization in the long run. Only when each component belongs to its own development, it is easy to optimize the performance. The separation of software and hardware will lead to performance optimization. < / P > < p > for comparison, NVIDIA Xavier has 9 billion transistors, Tesla FSD HW 3.0 has 6 billion transistors, but Xavier’s computational power index is not as good as hw3.0. And it is said that the next generation of FSD HW has seven times the performance improvement compared with the current. So it’s because Tesla’s chip designer Peter Bannon and his team are better than NVIDIA’s designers, or is it because Tesla’s methodology of combining hardware and software is better. I think the methodology of combining software and hardware is also an important reason for the improvement of chip performance. < / P > < p > for specific scenarios, we can imagine that NVIDIA can’t be as targeted as Tesla when designing Xavier. Their designers can’t know how the customer’s algorithm works, so they can only rely on the understanding of NVIDIA’s own algorithm and the conjecture of customer’s algorithm to design chips. < / P > < p > and Tesla’s chip design team certainly knows more about the algorithm, so it’s easy to plan the chip design. For example, convolution operation accounts for a large proportion, it focuses on optimizing convolution operation. According to the information disclosed by Tesla, it is clear that the proportion of convolution operation and relu is the proportion. Tesla’s internal team certainly knows more details than we do. < / P > < p > moreover, Tesla is not only responsible for chip Peter Bannon and Andrej karpathy, who is responsible for algorithms, can sit together to discuss. They will certainly find a way to measure the performance of the chip by using simulation or other methods before streaming, to find the bottleneck of performance, and then to optimize continuously. < / P > < p > for NVIDIA and their customers, there is no such close cooperation. Therefore, Tesla’s software and hardware integration optimization is better than the chip and algorithm disassembly in different companies. Therefore, in the field of automatic driving, it is not a good business to separate algorithms or chips and sell them separately in the long run. Even if it’s NVIDIA’s autopilot reasoning chip, of course, orin generation will still sell well, and the next generation may still be able to sell it, but it is likely that the more it will be sold, the worse it will be. Because of the performance and cost reasons, the major new power automobile manufacturers will eventually make their own chips. < / P > < p > of course, arm, acquired by NVIDIA, belongs to the mobile phone and terminal ecology, and has less to do with automatic driving, so it should have a good business. < / P > < p > then is it possible for a company to make all the automatic driving algorithm chip operating systems and provide them to the major car manufacturers. In other words, smart car = smart + car, one company does “smart”, other car companies do “car”, and then combine it into smart car. < / P > < p > secondly, in the face of consumers, the feedback is not as fast as that of Tesla and new force automobile manufacturers. If there is a traffic accident, will there be various partners shirking responsibility for each other. < / P > < p > in addition, aspiring auto manufacturers will not give up mastering algorithms and data, while unsuccessful auto manufacturers can give up their own algorithms, but they can’t do it because they are not striving for success. < / P > < p > so this Android route to build an intelligent platform is not the best solution in the long run. It is estimated that none of the new force car manufacturing enterprises that do well in the end will follow the Android route. This is different from that of smart phones. In the field of mobile phones, the performance and experience gap between Android solutions and apple solutions is small, while in the field of smart cars, the experience of closed-loop routes will be better than that of open routes. < / P > < p > of course, it is not necessary to achieve the closed-loop from soft to hard in one step. It should be realized step by step according to the needs of users and their own capabilities. Even Tesla is divided into several steps. < / P > < p > today’s automatic driving may only save 5% of the time, so only some consumers are willing to pay. As the proportion of this time increases, more consumers will be willing to buy smart cars. Therefore, the mode of self-made car sales can be gradually realized, invested and developed gradually. < / P > < p > and the robotaxi route aims to reach the goal in one step, because it is estimated that 100% automatic driving will be realized in at least 10 years or 20 years. In the meantime, robotaxi can’t really be commercialized.