This article is edited by CITIC's "Automotive" CPU "--from ADAS to Autopilot" and founder Securities "Automotive Semiconductor Research Framework"

With the increasing progress of self-driving technology and the improvement of ADAS penetration, the requirements for the number of sensors and computing power are also increasing, which directly stimulates the promotion and evolution of on-board AI chip technology.

Five Dimensions of comparison of vehicle-mounted AI chips

Which technology is the best in the car AI chip? In addition to the simple comparison of computing power, the energy efficiency ratio of each chip, the coupling of software and hardware, the openness of the platform, the product time planning, the realization of the chip function and the level of vehicle specification certification are all very important reference dimensions. Specifically:

Dimension 1: computing power and energy efficiency ratio

Computing power and energy efficiency ratio is the core competitiveness of chips. Chips with high computing power and high energy efficiency ratio can often achieve autopilot more efficiently and stably. In terms of computing power, the industry generally believes that L2 needs less computing power than 10TOPS, and L3 needs 30-60 TOPSMagi L4 to exceed 100 TOPSMagi L5 to exceed 1000TOPS. Chip is the core computing unit of self-driving car. In order to achieve high-level self-driving technology, the corresponding chip needs higher computing power.

At the same time, excellent chips not only need to be highly competitive in absolute computing power, but also take into account energy efficiency. Autopilot chips are often installed on electric vehicles, and the energy consumption of autopilot chips is also very important. The chip with high energy efficiency can not only save a lot of electricity for the car, but also generate less thermal energy, which is conducive to the heat dissipation and high performance and stable operation of the chip module.

In terms of numeracy, NVIDIA Corp's Orin chip is in the lead with the numeracy level of 200TOPS.Xavier chip is one of the most powerful chips in mass production on the market, and the company's solution is the most powerful. Although the EyeQ5 computing power of Mobileye is only 24TOPS, its energy efficiency ratio has reached 2. At the same time, the high coupling of hardware and software makes Mobileye show a good autopilot ability when the chip computing power is not high.

In terms of energy efficiency, Chinese companies Horizon and Black Sesame Technology plan to launch Expedition 5 and Huashan II A1000 reached 5 and 6TOPS/W respectively in the next two years, far ahead of their competitors.At the same time, although NVIDIA Corp's low energy efficiency has been criticized, its planned DRIVEAGXOrin solution in 2022 has an energy efficiency ratio of 2.7, and its computing power has reached 2000TOPS, which basically meets the requirements of L5 level and is very competitive.

Dimension 2: software and hardware coupling

The coupling degree of hardware and software of the autopilot chip has a very important impact on the efficiency of the chip, and the system with the combination of hardware and software often has better performance. Highly coupled hardware and software generally have higher computational efficiency, but it also limits the freedom of downstream manufacturers to develop in the field of autopilot. On the contrary, companies with low coupling between hardware and software can provide downstream manufacturers with higher degree of freedom and flexibility, but the disadvantage of this flexibility is that if there is no effective agreement between software and hardware, the performance of the chip may be greatly reduced, and it will also affect the stability of the driving system.

Take Mobileye as an example, its "software + chip" black box solution has very close software and hardware coupling and low openness.This feature enables them to achieve a very high level of autopilot even if they do not have an absolute advantage in chip computing power. But at the same time, their model is also controversial. In the era of autopilot, data is often the core competitiveness of a company, and no manufacturer is willing to cede the core data to the algorithm. Usually, Mobileye will only output the perceived target results after processing to the car company, but will not provide the original data. Therefore, many manufacturers will have more concerns when cooperating with Mobileye.

In contrast to Mobileye, NVIDIA Corp adopted a highly open solution.The NVIDIADRIVE platform developed by NVIDIA Corp provides a full range of customizable solutions from the underlying computing layer, operating system layer and application layer. Downstream companies can develop their own autopilot programs and own their own data on its platform with high degree of freedom. But at the same time, a high degree of freedom is likely to mean low coupling, the software algorithm research and development ability of traditional car companies is weak, and the landing performance of self-driving will be affected to a certain extent.

Except for these two companies.The horizon is positioned between the twoIt can provide open computing platform and software algorithm, thus improving the efficiency of cooperation with automobile companies, and differential competition has won more market favor.Black sesame is similar to NVIDIA CorpConcentrate on improving the performance of the hardware, and open the design of software and algorithms to the host factoryHuawei has its own research chip and development platform.To provide a complete solution to the outside world, the coupling of software and hardware is high, and the cooperation between car companies and Huawei is more cautious.

Dimension 3: product time planning

Foreign giants entered the game ahead of time, followed by domestic companies, speeding up mass production of chips to seize the market.The mass production time of chips largely determines the company's ability to seize the market, and the companies that are the first to launch qualified chips will always occupy more market share. therefore, various manufacturers are in the progress of accelerator chip design and mass production.

Foreign autopilot chip companies led by NVIDIA Corp and Mobileye layout ahead of time, has a clear advantage in the launch time, Mobileye plans to launch EyeQ5 chip in March 2021, to achieve L3 level of autopilot; NVIDIA Corp in 2019 mass production of Xavier chip, made great market gains, the company also plans to mass production of Orin chip one year ahead of schedule in 2022, cooperation with Li Auto Inc. to launch L4 level of autopilot.

At the same time, domestic chip manufacturers are also making great efforts to accelerate the progress of chip mass production. take Black Sesame Intelligent Technology as an example, the company has just launched the A500 chip at the end of 2019, and Huashan II A1000 in June this year and plans to mass production next year. The interval between the two products is less than 300 days, which shows the efficiency of research and development.

Dimension 4:ADAS function realization degree

The evolution of chip generation often means the realization of higher-level ADAS functions, and the ultimate goal is to achieve L5 fully autopilot. ADAS provides assistance and supplement for drivers in the complex process of vehicle operation, and finally realizes self-driving, including FCW, LDW, LKA, AEB, ACC, ICC, TJA, APA, TJP, HWP (HighwayPilot, highway autopilot), HWC, CCF (urban autopilot) and so on.

From the evolution of the chip, we can see that low-level chips can only achieve L1 and L2-level auxiliary driving, and with the improvement of chip computing power and the maturity of technology, companies can gradually achieve a higher level of assisted driving. At present, Mobileye's EyeQ5 and NVIDIA Corp's Xavier can achieve the most functions and perform best in the same generation of chips.

Dimension 5: vehicle specification level certification

Passing the specification-level certification means that the chip can enter the market faster, and the higher the certification level, the more popular it will be from the whole car factory. Automotive standards need to be certified as one of the reliability standards AEC-Q series, quality management standards ISO/TS16949, and functional safety standards ISO26262ASILB (D).

ISO26262 mainly consists of four levels, namely ASILA/B/C/D. ISO26262 safety is one of the criteria for judging the stability of automotive electronic components, which means that the stability of the product is qualified and durable, but it does not represent high computing power and energy efficiency.

The standard of vehicle specification grade chip is much higher than that of consumer grade, and the certification process is long, so the certified chip often has stronger market competitiveness. The working environment of automotive chips is even harsher. Compared with consumer chips and general industrial chips, automotive chips have a wide temperature range (- 40 to 155 degrees Celsius), high vibration, dust and electromagnetic interference.

At the same time, cars have high requirements for the reliability and safety of chips, and the general automobile design life is about 15 years or 200000 kilometers, which is far greater than the life requirements of consumer electronic products. Under the same reliability requirements, the more components and links the system consists of, the higher the reliability requirements of the components.

In addition, the certification process of the on-board chip is also long, a chip generally takes about 2 years to complete the vehicle specification level certification, and generally has a supply cycle of 5-10 years after entering the supply chain of the automobile company. At present, EyeQ5 and Orin have the highest certification level in ISO26262, both reaching the ASIL-D level; domestic chips perform well, and most of them have passed two certifications.