Text | Karakush
When the industrial revolution comes, you don't know who is going to squat at the door next.
After last year, many car companies must have a deep understanding of this. The chip problem can not be solved, so that car companies began to directly collude with chip factories, some special "guarantee and supply office" to squat, and some bosses personally came to the door to squat. What was originally a sandwich relationship between Tier 1 or several layers of other characters became a sour melon patch slice, back-to-back face intimacy.
The car family is unfortunate to "chain" the family lucky. Supply chain security is not so safe, and the sense of crisis that car companies have suddenly emerged is not only aimed at the old ass that is currently out of stock, but also extends to vigilance against the future of autonomous driving chips. This has paved the way for many specialized new chip factories to get rich.
Two days ago, we chatted with a local self-driving chip company Black Sesame Intelligence, and found that those grasshoppers who were originally difficult to shake the tree were waiting for an opportunity to get on the car. Standing at the industrial level, you can no longer underestimate the small amount of fruit juice and flesh of the locust ant juice, and the things that are developing must be treated with the vision of development, and the empires of tomorrow that you are staring at or will be just around the corner.
In the past decade, the most successful transformation of China's auto industry is not BYD or Wei Xiaoli, but a trillion-level supply chain enterprise like the Ningde era, and its momentum has surpassed any new energy vehicle company. As intelligent driving matures towards commercialization, do we have the opportunity to grow more Ningde in the future, such as in the field of chips, there is also a representative and leading enterprise in the world?
Today this is something that can be boldly imagined.
A ticket for the White Era
The biggest impression of black sesame intelligence from the outside world, if any, is probably that the Xiaomi Yangtze River Industry Fund led its Series C financing in September last year. Black sesame seeds became a footnote to the millet-made coarse grain universe.
In fact, at that time, black sesame had already achieved great results, and released the A1000 chip in 2020, which was the earliest large computing chip in China, with a single 58~116TOPS, one year ahead of its peers; last year, it released A1000Pro, and the single one reached 106TOPS (INT8)/196TOPS (INT4), which belonged to the strongest domestic and international first-class level.
You may have seen the earliest and biggest formulations elsewhere in the press release. The measure of black sesame is the success of the tape-out (that is, trial production), according to the car company can get the hand of the large-scale automatic driving chip, "at present we (black sesame) is the only one." This year, they will gradually mass-produce on the car, and the specific model information needs to wait until the official announcement of the cooperative car company.
Still, it's as famous as it is. Founded five years ago, the first three years have focused on self-developed chip core IP. Compared with the purchase of general IP, the advantage is performance improvement, especially some custom development related to product functions, and the downstream product ecology can establish a close coupling.
With the explanation of human Chinese, the current two core IPs, one is the ISP digital image processing unit, which can make the image data collected by the sensor achieve a more uniform effect under different light changes. For example, when the vehicle enters and exits the tunnel, it will go through the process of dawn to dark, dark to dawn, shadows or glare will sometimes flash our titanium alloy human eyes, and the chip will also be affected by the same, to do some front-end processing through the ISP, the back-end algorithm can play a better effect.
The other is the NPU Deep Neural Network Acceleration Unit. Cars need to identify and analyze various environmental data such as vehicles, pedestrians, lane lines, traffic lights, etc. collected by sensors, and abstract the environment into maps and obstacles in order to understand the world and further make automatic driving functions. NPU allows engineers to run efficiently based on computer-based models in the R&D environment directly on the car.
The above is to illustrate that this is a company with two brushes on the product and technology. And can do in line with the car rules - in the market thousands of chip design companies, only 1% can do the car rules level.
In the automotive industry, however, talent is only a necessary and insufficient element of success.
In the past, Chinese chip design companies mostly bypassed cars. One is that there is no return, the difficulty is large, the cycle is long, to make a chip, from the beginning of the definition to the final appearance of the car, it takes 5 years, 5 years will not generate income; the second is that it is difficult to find an export, a wild start-up company is almost impossible to enter the traditional car company's relatively closed supplier system.
Now, a new turnaround and a bucket of stars are filling the slightly open sky.
The first is the transformation of automobiles into electronics. There are about four or five hundred chips on traditional fuel vehicles, and there are generally more than a thousand smart electric vehicles in this generation, and even three thousand high-end models. The increase does not come from the existing dozens of chip violent replication, but the emergence of many new chip needs, providing opportunities for new product lines for startups.
Secondly, the supply chain crisis has prompted car companies to attach importance to the establishment of a local supplier system, pay more attention to the source of chips, supply quality, maintainability and sustainability, and open the door for new suppliers to enter.
Three years earlier, no possibility. Autonomous driving experienced an industry trough in 2018 and took off again in the second half of 2020. At that time, Black Sesame just launched the first large computing chip, which was accepted by car companies after a year, and then prepared for mass production after a year, the rhythm was precise, and this year's launch has a time advantage.
Three years late, it is not possible. The best time window is only three to five years, after which it gradually closes. No matter how the technology is iterated, the automotive industry itself will remain relatively conservative. This is different from consumer electronics, where the supplier generation did not catch up, and eight months later the next generation opportunity came again. After 2025, the new supplier system of car companies will be established intact, and there will not be many opportunities for new companies behind.
"We're in the best of times we can do this."
For this batch of chip companies in China, they all need to quickly establish their irreplaceable position by taking more car companies, more models, iterating products and mature technologies.
At present, the primary consideration of autonomous driving chips in the industry is foreign manufacturers such as NVIDIA and Qualcomm. In particular, NVIDIA, in addition to the leading and mature products, also provides an intelligent brand endorsement for car companies, buying one and getting one free so that peers can't look forward to it.
In contrast, the share of domestic chips is extremely low, and the statement of high emotional intelligence is that "the future domestic head growth space is huge" - the fact that there is indeed a chance. The field of autonomous driving chips is a deep blue ocean, the market space is abundant, due to the relatively high threshold, the players can be predictably few, and the provision of large computing power chips and service spirit is even fewer.
This becomes an advantage for domestic startups.
Use cheapness not to pick work, to move China
The attitude of domestic startups is mostly correct, cheap and not picky, and the official language is "with a flexible business model", soft enough and hard enough. For example, in order to facilitate customers to get started, black sesame also provides full-stack algorithms in addition to chips; if customers have other algorithm resources, they can also do nested fusion.
In fact, from the stage of new product definition, they will give priority to soliciting the next-generation function points and performance points of domestic car companies, in order to adjust and customize chips and systems, and the rhythm of technological evolution is close to the needs of car companies, and every crystal is shouting services.
The real difference compared to attitudes is the mode of cooperation. When foreign chip factories are still contacting car companies through Tier 1, these domestic startups are trying to directly connect with car companies. Personal service, not only the response to demand and feedback will be faster, the cooperation model itself reflects that in the process of transitioning from traditional cars to smart cars, the industrial structure has changed, the relationship between chip factories and car companies will be more entangled, and the right to speak and importance will gradually be heavy.
In the traditional supply chain structure, chip factories belong to Tier 2, and car companies generally only dock to Tier 1, and the two are not related. For China's independent car companies, the early research and development capabilities or the ability to raise demand are relatively weak, mention the demand to Tier 1, Tier 1 packaging technology to provide a black box, car companies to pay a higher cost to buy it. Therefore, under the original pattern, car companies control products, and technology is promoted in Tier 1.
To this generation of smart cars, more and more car companies have begun to assume the role of leading the direction of technology. Whether it is a new force or a traditional car company that strives to transform, it is expanding the ability of intelligence and electrification, and wants to participate more in the formulation of technical routes, master their own technical points, and make differentiation, which will inevitably bring about in-depth research within the system, such as actively investing in some early POC and pre-research, to learn quickly more upstream.
Of course, these include automatic driving chips, many times, the chip does not support the function, the terminal can not play tricks, car companies attach great importance to this part that determines the future direction. Some car companies have then gone to the other extreme - self-developed chips, especially at a certain point in time in the previous year, a number of car companies have been exposed to self-research plans.
Apple and Tesla are the model of Niu B, but the self-development of full-stack closed systems is difficult to replicate in each industry, which is determined by the stage of industrial development. Whether it is Apple or Tesla, when choosing self-research, it is not supported by the industrial chain in the early stage. When they successfully define the technical solution and drive the entire industry, the latecomers who want to do something similar will not need to invent the wheel from scratch.
Now chip manufacturers and startups are engaged in autonomous driving chips, compared with car companies, more focused on the field, and can be achieved at a faster speed and lower cost. From the perspective of business logic, the first 60% of the things in the chip are universal, and the differentiation is 30 to 40%, so the most cost-effective way is to find a chip company to customize.
Of course, this is the enterprise-level understanding of chip factories. Now there are still car companies insisting on going it alone, and will also find black sesame for technical exchanges, to which black sesame said, "Customers hope, we must all feel good." ”
As for the mainstream relationship between chip factories and car companies in the future, it is expected to develop from the traditional OEM-Tier 1-Tier 2 to OEM-Tier 1.5. In the past, chip factories supplied Tier 1 in series, and now they will dock with car companies one by one to see what is needed, and then cooperate with Tier 1 to develop and provide more complete solutions.
In the future, car companies will gradually differentiate into various forms. One class focuses on customer needs and experience, does product and brand definition, and relies more on Tier 1 for technical solutions; there will also be a class that takes one step further, assumes the role of Tier 0.5, and works with the industrial chain to create products with customized characteristics.
China's hot soil is enough to take the lead in completing such a transformation. The intelligence of China's auto market is very advanced, and only Chinese car companies are keen to load 1000T and can roll it up. Its degree of radicalization will also play a pulling role in improving the upstream chip capabilities of local industries.
Although these 1000Ts are all NVIDIA customers.
How to make autonomous chips stronger
In the future, the biggest core competitiveness of chip factories will still fall on the demand for car companies and Tier 1. However, before looking at 100, the current 0-1 market competition is still oriented by hardware performance, and the art of pinching is not enough to show strength.
How much computing power do we really need, and what kind of hardware performance is the end of the universe? There are two observation dimensions.
From the inside out, as a selling point, at least flagship models or high-end models need to be standard or even over-equipped, it is an important reason to attract users to buy, symbolizing the degree of car intelligence, even if the two can not be said to be unrelated, it is not exactly the same. In the existing stage of consumer cognition, the greater the computing power, the better.
From the outside to the inside, now the car intelligent driving is mainly in L2 ~ L2 +, the computing power requirements are not high, a few LARGE TOPS, up to 100 TOPS is enough. However, automakers need to embed the computing power of the software that is constantly upgrading and the experience that is constantly improving, and through the accumulation function of OTA, more computing power will be used. So in any case, big is good.
For any electronic products, from PCs to mobile phones, in the early stage of development of hardware performance is an inevitable law, chip performance first rapid iterative improvement, in order to provide a stronger functional support foundation for software and applications, in order to develop software and applications; upgraded to the time when it will not become a bottleneck in development, selling points and innovation points will gradually move to software and applications.
The same goes for cars. In the years when people have been shouting about software-defined cars, hardware has actually made a qualitative leap. The computing power of the chip has changed from tesla using Mobileye only 0.256TOPS six years ago, a level that now seems to belong to the smart floor; to the more popular Snapdragon 820 2.5TOPS; to 2019 Tesla's self-developed HW3.0 to provide 144TOPS; to Xiaopeng using NVIDIA Xavier to provide 30TOPS; to this year Nvidia Orin will mass-produce 1000+ TOPS on Weilai and Zhiji.
The software experience is nowhere near as visible as hardware data grows. The process is not finished by the car.
High-performance computing chips will be the core of future electrical and electronic architectures. At present, most cars are still using distributed ECU architecture, gradually shifting to domain control architecture, and finally to central computing architecture, and the higher the chip performance requirements in the future. Under the central computing architecture, all computing will be attributed to a chip, with the improvement of technology and process, there is a chance to achieve, many manufacturers are mentioning.
On the other hand, the competition for computing power will not grow indefinitely. There is a view that 1000 ~ 2000 TOPS will be the peak of computing power growth in the short term, just like ordinary cars to four wheels capping, will not go to five rounds and six rounds, and now four chips are a great challenge to energy consumption. How to reduce energy consumption, improve energy efficiency ratio, safety level and other parameters will become a new demand.
From the perspective of reality, these are not only the explorations that the industry is doing, but also the afterword. If the 0-1 process is subdivided into chip samples, customer fixed points, and mass production on the car, all domestic large computing power chips are between two and three steps. This year is crucial.
For these chip factories, peers are not rivals. In an incremental market, competition is not fierce, and it is said that peers often sit down to eat and exchange their progress.
It is more important to prove yourself as soon as possible, run through the scale, complete the market verification, and have a good experience, at least not collapse. The car companies that originally held the currency and waited to see if other families were happy, it was easy to follow up, the car companies followed the trend, and the wind called out to the insider with tears. Start with local companies to increase their share; it seems that the fierce competition is mainly public relations.
Only by virtue of China's intelligent automobile industry, wanshan is red, layers of forests are dyed, and there is still more than enough to support one or two world-class automotive chip companies.