Only after a thousand entreaties does she appear,$NVIDIA (NVDA.US)$The long-rumored Thor Hammer has finally been officially released. At the GTC conference in 2022, NVIDIA Corp cancelled the Atlan autopilot chip released at last year's GTC conference and replaced it with a more powerful Thor (Raytheon). Its performance supports FP8 format, and the computing power of single chip AI can reach 2000TFLOPS, which is twice as high as that of Atlan. Functionally, it can dominate the cockpit domain and intelligent driving domain. There has been a lot of introduction to other parameters of Thor in various media, so I won't repeat them here.

To sum up, you only need to understand four points:

First, the computing power in the field of autopilot has entered a new prospect stage.

Second, Thor is the most powerful autopilot chip at present.

Third, the competition between NVIDIA Corp and Qualcomm Inc in the field of autopilot has escalated.

Fourth, the threshold of the technical ability of autopilot has been raised again.

As soon as Thor was released, it aroused a burst of excitement in the self-driving industry, as if self-driving was just around the corner, and even a large number of "English blows" emerged. But calm down and reflect.NVIDIA Corp has only taken a step forward in his business landscape and cannot be the decider of the development of the self-driving industry.

The ups and downs of the self-driving industry has once again confirmed the periodicity of the application of new technologies, and every progress must be on the shoulders of our predecessors. From the release of 1TOPS's Parker in 2016 to the release of 2000TOPS's Thor, NVIDIA Corp has updated Xavier (30TOPS) and Orin (256TOPS) in just six years, of which the mass-produced model of Xavier is the 2020 XPeng Inc. P7, while the mass-produced Orin in 2022 has been announced by many car companies or self-driving solution companies, such as NIO Inc., SAIC Feifan, ideal, Wenyuan Zhixing and so on. From the point of view of the time node, most car companies are developing models in mass production based on Orin films, while others are still waiting and testing.

This year's GTC conference NVIDIA Corp released a stronger Thor, and plans to carry a specific brand (polar krypton) model mass production in three years' time. That is to say,The Orin-based scheme can be used for at least another three yearsAs for whether the car companies choose to continue ploughing Orin or chasing the tide is the same effect for NVIDIA Corp. So on the other hand, it seems that NVIDIA Corp is the real demand side of the product, has been guiding the intelligent development of the automotive industry, and car companies or self-driving solution vendors have become their "tool people". In other words, no matter how powerful the autopilot chip is (the physical limit of the process is difficult to improve qualitatively)NVIDIA Corp always has more advanced hardware waiting for the market, while car companies always have endless threshold fees.. It seems that the car company is like a donkey in a circle, thinking that it is working for itself, but in fact it has been working for the landlord.

The impact of Raytheon's release can be seen in several ways:

First, the impact on the domestic chip industry.

Self-driving in the industrial innovation of hardware first, domestic and foreign chip manufacturers can be said to rise together, big computing chip has become standard, the development speed of chip computing is obviously faster than the application side of self-driving. With the Thor sacrificed by NVIDIA Corp, the pursuit of computing power by domestic chip manufacturers will be more enthusiastic, and the transmission to the application side is that the autopilot industry will start a new round of arithmetic war. The domain control based on Raytheon will integrate the functions of car machine, instrument, electronic rearview mirror, DMS and so on, so it will be more difficult for domestic AI chips to develop the market. Furthermore, Raytheon was NVIDIA Corp's first chip to support the FP8 format. As we all know, NVIDIA Corp is far ahead of Huawei, Alphabet Inc-CL C and other competitors in the training chip.Raytheon tries to unify AI training and reasoning, so the living space of domestic AI chip enterprises that do training or reasoning alone is further compressed.

Second, the impact on the automobile industry.

The first point is that embedded systems are essentially developed under the condition of limited resources.Different from the server development mode of the Internet, this resource refers to the performance of the chip, time node, product demand, product target cost and so on. Whether it is cockpit system or autopilot system, the trend of domain integration is actually the deep integration of computing power, power consumption, time and other resources. The sharp increase in chip computing power seems to be a war of affluence, but from a commercial point of view, it is very difficult to work. At present, most domestic chip companies have conducted in-depth cooperation with car companies at the capital level or product level. Raytheon chip car companies will face a balanced choice of capital or technology.

Second, trigger a new round of changes in autopilot technology.This change is not only in terms of chip computing power, power consumption and computing efficiency, but also in terms of software and data and even development patterns. For example, on the Raytheon film, the special engine prepared for the recent hot large model Transformer in the field of vision, based on the advantages of hardware platform, more enterprises will join the pre-perception fusion route (see Tesla, Inc.).

The third point is to provide a new thinking direction for automobile intelligence.The intelligence of the automobile makes the system more complex. Whether the distributed architecture or the centralized architecture is a process from divergence to centralization in thinking, the difficulty is that human beings are good at the scope of divergent problems and lack of convergence. From auxiliary driving to advanced assistive driving to self-driving, the industry has been trying to open up this development path. The reality is that the problem of disconnection will occur no matter from high to low or from low to high. The algorithm has been in a state of optimization, and the limits of data and computing power have never been reached. The emergence of big arithmetic chips such as Raytheon provides new possibilities for exploring the limits of arithmetic. However, it is important to pay attention to the limitation of resources.

Fourth, car companies or autopilot operators are caught in a "prisoner's dilemma".NVIDIA Corp will promote the single-chip computing power to 2000TOPS. Car companies or self-driving solutions will think that the size of computing power determines the key cost of autopilot, the degree of intelligence determines the degree of intelligence, and the power of calculation determines the user experience and even the advanced nature of the product. First of all, it is bound to cause the internal volume of computing power in the car, and secondly, the communication between enterprises will be reduced. People prefer to work behind closed doors. If the product is not ideal, continue to add data and computing power. Cycle over and over again. There is only one solution. To design products based on practical application, not all problems can be solved through data and computing.

The last part is the influence on NVIDIA Corp.

First of all, we have to admit that NVIDIA Corp is a technology company with geek spirit.

You can not accept your opponent, but you can't help but admit the greatness of your opponent. The annual GTC conference NVIDIA Corp will bring new products, Raytheon's release to some extent highlights the continuous pursuit of semiconductor performance. There has been a discussion in the industry for a long time as to whether NVIDIA Corp is a hardware company or a software company. With the product release in recent years, at least in the automotive field, the answer has been relatively clear.

Since the launch of the autopilot platform NVIDIA DRIVE on CES in 2017, NVIDIA Corp's software ecology and tool chain have been constantly improving. The NVIDIA DRIVE platform supports taking the DRIVE AGX development platform as the starting point, verifying the software algorithm on DRIVE Constellation, fully verifying and deploying the software, then testing on the road through the DRIVE Hyperion reference architecture, and finally using NVIDIA DGX for artificial intelligence DNN training, so as to reach the mass production standard. In the process, the platform also provides developers with a rich software developer suite (SDK). Generally speaking, people pay attention to the release of more specific chips. in fact, NVIDIA Corp has a big chess game under such a fast iterative speed of the product.That is the tool chain and the continuous improvement of CUDA-based ecology.. After the rapid iteration of Parker, Xavier and Orin products, NVIDIA Corp has formed a complete tool chain and software ecological force, which is very helpful to improve the development efficiency. In this case, not only the chip peers have the pressure of product competition, but also the car companies have certain choice pressure in the face of cost and technical capability.

Secondly, NVIDIA Corp's life this year is not easy.

First, the mainstay business game graphics cards suffered an avalanche drop of 30%. Some time ago, graphics cards began to be sold at reduced prices, which was unthinkable a year or two ago.

The second is that the chip bill of the United States passed the chip bill. The drunken man's intention is not to give away money and subsidies, but to make semiconductor manufacturers stand in line to strengthen their political voice on science and technology.

The third is that the United States restricts NVIDIA Corp's export of high-end graphics card chips A100 and H100, while the reality is that NVIDIA Corp has about 10% of data center chip sales in China.

Fourth, NVIDIA Corp's share price has halved compared with last year, and the capital market began to vote with its feet.

Although NVIDIA Corp seems to take the lead in the automotive field, in fact, NVIDIA Corp's "surplus grain" is also gradually decreasing. China is not only an important market for NVIDIA Corp graphics card sales, but also a hot land for automobile intelligence. so Raytheon's first cooperation or most of its partners will also be in China, waiting to see which brand will be the second. From the perspective of Tongying Weida's overall business, Raytheon is also a helpless choice for NVIDIA Corp to adjust his strategy.Although domain integration is a trend, the demand for computing power can not be quantified because of the uncertainty of product demand.In this case, the stability of Raytheon's application is almost certain.

Each industry has its own unique industry attributes, each company has its own development problems, the industry and enterprises are not wrong, every transistor and every line of code will promote the progress of automobile intelligence. With the direction and the choice, the rest of the test is the product development strategy. If you choose too much, you will be lame if you take too many steps. Blindly pursue the wind mouth to engage in publicity, Orin has not yet studied and understood, began to plan Raytheon, in the case of the same resources, the result will be lame.

This article is edited from the official account of "Smart car Science and Technology" Wechat, author: Shi Fu, Zhitong Financial Editor: Yang Wanlin.