Text/Car Bright
Graph/from network
Zhou Hongyi, the founder of 360 Company, once said: "In the past, we measured whether a car was a luxury car, horsepower was a very important indicator, and when the automotive industry was subverted by the network and digitalization, the standard for measuring luxury cars in the future was not horsepower, but computing power."
Li Bin, the founder of Weilai, also said that the standard of high-end intelligent electric vehicles in the future is horsepower + computing power.
The concept of computing power is now being paid attention to and valued by more and more consumers. For example, the recent scandal of Euler's good cat, the official website said that qualcomm Snapdragon's 8155 chip was used, but in fact, it used an older Intel chip, which caused consumer dissatisfaction. Although the official compensation policy has been given, many consumers still say that they would rather have Qualcomm's chips than compensate for the gift package. Behind this psychology of consumers, it is actually a pursuit and desire for high computing power.
So is hash rate really so magical? What exactly does hash rate mean for a car? Today we will talk about the topic of computing power.
Computing power is actually productivity
What is hash rate? We have to start with the three elements of artificial intelligence. The so-called three elements refer to data, computing power, and algorithms, known as the troika of the digital age.
Data can be understood as means of production, and whether the data is massive and rich is the premise of all intelligence. So now some car companies are playing the concept of lidar, are frantically stacking cameras, millimeter wave radar, etc., just to collect enough data. Data is the source of everything, and the so-called smart women are difficult to cook without rice, which is the truth.
But data alone is not enough, there must be a reasonable algorithm. Algorithms can be understood as production relations, which are rules and methods for processing data. The function of the algorithm is to draw a law based on massive data, and then use this law to model and predict the unknown data. Simply put, it is to transform all the real problems encountered in the car into mathematical models.
In the previous car, the various parts of the control unit ECU is a separate layout, the algorithm of each part is relatively simple, similar to an independent small workshop, once the basic structure of the algorithm is determined, that is, the production method, it may not change. But now, with the transformation of the electrical and electronic architecture on the car from distributed to centralized, the various control units are concentrated together, which is equivalent to hundreds of small workshops being integrated into a large company, and the production method and management difficulty are greatly improved. Coupled with the rapid development of smart cockpit and autonomous driving technology, it has brought a large number of application scenarios and a large number of random and unstructured data, such as face recognition, voice control, and gesture recognition, how to deal with these data? For example, a large number of sensors have been added to the car, how to deal with the data transmitted? These require new algorithms. So as cars become more intelligent, the requirements for algorithms are also rising. For example, the newly added algorithms in the intelligent driving part include perception algorithms, multi-sensor fusion algorithms, decision-making algorithms, control algorithms, etc., and the newly added algorithms in the intelligent cockpit part include speech recognition, face recognition, gesture recognition, line of sight monitoring, scene understanding and so on. So many algorithmic models have never appeared on previous cars.
The algorithm corresponds to the production relations, and the data corresponds to the means of production. Then the computing power corresponds to the productivity. The so-called computing power refers to the computing power provided by basic data processing and algorithm training with high-performance computers as the carrier. In addition to the processing of data, another important role of hash rate is to support the training and formation of algorithms. Just as productivity determines the relations of production. The training and formation of algorithms also depends on the level of computing power. For example, algorithms are like ways and means to solve problems, and computing power is like the ability to solve problems. Only when a person has the ability can he come up with a solution. And a person has the right way to give full play to his ability.
Therefore, from the three elements of means of production, productivity and production relations, we can easily understand the relationship between data, computing power and algorithms.
Understanding this layer of relationship, we can see more clearly and clearly the propaganda of some car companies advocating high computing power and advocating hardware.
For example, some car companies advertise how much lidar and how many cameras are used in their cars. But more hardware does not mean that your effect is good. The data you collect is massive, is your hash rate keeping up? Is your algorithm model reasonable? The latter two are actually more decisive about how you end up. And now many electric vehicles are taking the approach of hardware embedding and slowly developing software in the later stage. That is to say, after the consumer buys a car, he has to wait for an unknown amount of time before he may use these functions mentioned in the promotion. Why? That's why the algorithm didn't keep up. Like Tesla's autopilot, insisting on using cameras and not using lidar, the computing power of the chip is not too high, but the final effect is good. This is also due to the strength of Tesla's own algorithms.
For example, some car companies publicize how many high-computing chips they use, and the basic computing power is as high as thousands of TOPS, etc., and the final effect must be good? not necessarily. Just as the productive forces determine the relations of production, the relations of production in turn promote or constrain the productive forces. Algorithms also promote or restrict the use of computing power. Therefore, high computing power will not necessarily have good results. It also depends on whether your algorithm model is accurate and matches. Therefore, the relationship between computing power and algorithm is also a process of continuous mutual promotion and spiral.
So, data is the source of everything, and nothing can be said without it. Computing power is the cornerstone of everything, it determines the efficiency of data processing, and it also drives the continuous training and evolution of algorithm models. But the algorithm is the key to determining the final effect, which determines the lower limit of whether the data and computing power can play their due role.
For car companies, electric drives, batteries, lidar, and chips can all be mined, and the hardware may not be necessary to do it themselves. But software algorithms must be firmly in their own hands in the future, which is the key to victory. Just as the Volkswagen Group claims to transform itself into an "automotive software company" in the future, mastering software is the core competitiveness of car companies in the future.
What level of computing power is really high?
The level of computing power has now become a point for many car companies to publicize to the outside world. In fact, you will find that the level of computing power promoted by car companies is actually very uneven, there are only a few TOPS, there are hundreds of TOPS, then in the current market environment, what kind of computing power is really high?
The carrier on which the computing power is based is the chip. In fact, when it comes to the computing power of the on-board chip, it is divided into two sections: intelligent cockpit and intelligent driving.
In the field of smart cockpits, the most dominant in China is Qualcomm in the United States. At present, the most advanced production car is its third-generation intelligent cockpit chip, that is, the 8195 chip. Currently equipped with this chip are Cadillac's LYRIQ and the upcoming NIO ET7. Also belonging to Qualcomm's third-generation chip is the 8155, and the representative models equipped with this chip are Xiaopeng P5, SAIC Zhiji L7, Great Wall Mocha, Zero Run C11, WM W6 and so on. Further down, the most mainstream currently used is actually Qualcomm's 820A chip, such as Weilai, Xiaopeng P7, Audi A4L, Lynk & Co 05, Extreme Kr 001 and so on are equipped with this chip. In addition, the smart cockpit chips of domestic manufacturers also have some market share, like the Horizon Journey 2 chip, which is mounted on Changan's UNI-K.
The most advanced smart cockpit chip known now is from the recent Baidu-owned Jidu Automobile, which claims to use the Qualcomm 8295 chip, which is known as Qualcomm's fourth-generation automotive digital cockpit chip. However, it will not be unveiled until next year's Beijing Auto Show, and it will not be available until the following year.
Let's look at the autonomous driving part. In fact, the requirements for the computing power of the chip in the cockpit part are not too high. The real high demand for computing power is the automatic driving part. Smart driving chips are basically mobileye and NVIDIA two worlds at present, qualcomm is eager to try to three points of the world. At the same time, Huawei HiSilicon, Horizon, Black Sesame and other independent brands occupy a place. Looking at the computing power of mass production vehicles in the field of autonomous driving, you will find that their levels vary from a few TOPS to a few hundred TOPS. Like NVIDIA this year released up to 1000 TOPS chips, and NVIDIA's current chip has reached a maximum of 254 TOPS, Huawei's MDC810 intelligent driving computing platform, the computing power has reached 400 TOPS.
At the same time, there is also a trend now, because the car is basically a complete replacement in 5-8 years, and the cycle is very long. The iteration of the chip is one to two years. Therefore, in order to make their products have continuous competitiveness, car companies will adopt hardware embedding and software replacement to maintain competitiveness. Simply put, my current chip computing power is OK, and I may be outdated in two years. In order to remain competitive, I installed a few more chips, increased the upper limit of computing power, and provided enough room for subsequent software optimization and algorithm upgrades. So now some new forces have started an arms race to increase the computing power of the chip on the car to the level of 500-1000TOPS. Like Weilai's ET7, it is equipped with four NVIDIA chips, and the final force has reached 1016TOPS. Like Xiaopeng's G9, it is equipped with two NVIDIA chips, and the final force has reached 508TOPS. Is it possible to use such a high computing power now? It is not practical, but I install it first, on the one hand, it is a highlight of publicity, on the other hand, it may be used later.
Therefore, for consumers, the higher the computing power, of course, the higher. But don't mythologize the role of high computing power, because of the constraints of the algorithm, a lot of computing power is actually wasted at the moment, it may only be prepared for the future scene, you are just paying for the future.
Driving value view
Now chip manufacturers generally do not sell chips alone, but sell a complete set of intelligent platforms, that is, chips + algorithms packaged and sold. Car companies generally have to adjust and modify the algorithm according to their own needs. In other words, car companies must have their own ability to develop algorithms, or car companies must at least have the ability to put forward algorithm improvement needs. This ability determines how effective your final thing will be. This ability is often invisible in the publicity of car companies, and can only be perceived by consumers in the final product.
Computing power is only a number, we still have to look at how to make the algorithm software run the most efficiently under the limited computing power, can perfectly adapt to the needs of car companies, which is a test for chip manufacturers and car companies. This is the only criterion for the ultimate. Therefore, companies like NVIDIA, Huawei, and Horizon, the algorithm ecology is open to car companies, and car companies can improve and customize according to their needs, and such companies will be favored by more and more OEMs. Companies like Mobileye, on the other hand, have an algorithmic ecosystem that is closed and cannot be modified, so it is going downhill.
The intelligent experience of a car is ultimately a systematic engineering of data, computing power, and algorithm interaction. The level of computing power does not directly determine the quality of the final experience. Installing lidar to perceive data, using large chips to increase the upper limit of computing power, these car companies advertise things, is not the core capability, because these can be spent to buy. Some car companies blindly emphasize and show off how high their computing power is and how much lidar there is, which is actually an act of fooling people. For car companies, what are the core capabilities? If you don't do hardware and don't develop chips, algorithms are the only core capabilities you can have.
END