Car stuff (public number: chedongxi)
The author | James
Edit the | Xiao Han
The field of autonomous driving chips has entered a new player, this time it is Ambar.
Just last week, In the first week of the opening year of 2022, Ambarco launched the CV3 series of AI domain controller chips based on the 5nm process, which can reach 500eTOPS with AI equivalent computing power and can be used for L2+ to L4 autonomous driving.
In today's increasingly fierce competition in the autonomous driving industry, Ambarco's autonomous driving chip achieves stronger computing power while also has a higher energy efficiency ratio. Products with both high performance and high performance are of greater significance to the autonomous driving industry.
After the Amba CV3 series was unveiled at CES, an online launch event was also held in China this week. After the release, Feng Yutao, general manager of Amba China, and Xi Jianjun, vice president of marketing of Amba China, were interviewed by many domestic media such as Chedongxi, revealing more details of the CV3 series of chips and the views of the autonomous driving industry.
As a semiconductor listed company, Ambarco has made great achievements in the field of CV chips for many years, and the performance of autonomous driving AI chip products is more powerful, coupled with multiple mass production results from 2020 to 2021, the autopilot chip market is becoming more lively.
So, how well does the CV3 series chips perform? How does Ambarra do it? After a dialogue between Che Dong and Amba, the two executives found the answer to the question.
First, it has a maximum of 16 core CPU equivalent computing power 500eTOPS
Amba unveiled at this CES is the flagship product of the CV3 series SoC, which has reached the highest level of AMBAR in cpu core, neural network engine, general-purpose vector processor and other configurations.
The CV3 series chips are built using a 5nm process, which has a CPU with up to 16 core ARM Cortex-A78AE cores, which are divided into 4 clusters, each with 4 cores.
▲CV3 architecture diagram
Feng Yutao introduced that the reason why so many CPU cores and clusters are used is to meet the needs of different car companies. He pointed out that the CV3 chip is put into 4 clusters, which is designed for L4 automatic driving, and different clusters are used for different calculations.
▲Feng Yutao, general manager of Anba China
For example, you can use a cluster as a security island, a cluster for neural networks and visual perception, a cluster for fusion, and redundancy considerations. In this way, clusters of different cores are used as unused functions, and the software development for chips will be relatively independent, accelerating the progress of R&D and mass production of car companies.
According to reports, the computing power of the 16-core CPU can reach 270K DMIPS, and the performance performance is very strong in similar products.
For L2+~L4 level automatic driving, the neural network engine is naturally one of the most important cores, and the equivalent computing power of CV3's neural network vector processor (NVP) can reach 500eTOPS.
Feng Yutao introduced that e in eTOPS represents equivalent (equivalent), and the reason why the unit of counting THE AI computing power of the CV3 chip is different from most autonomous driving chips is because its architecture is more special and not equivalent to the GPU.
Therefore, the computing performance of the CV3 chip NVP core is difficult to compare with other GPUs. In the early test, Ambarco's engineers used CV3 and traditional GPUs to run the same algorithm, if the two are equally fast, then the equivalent computing power of CV3 is the same as the GPU of the comparative test.
He also said that in fact, Ambarco has already applied the concept of eTOPS in the previous generation of product CV2, so car companies customers already know how much its performance can be played. The CV3 has 42 times more AI performance than CV2, which is a huge improvement.
In addition to the NVP, a general-purpose vector processor (GVP) is integrated in the CV3, which is capable of achieving 920GOPS performance. Feng Yutao introduced that NVP is mainly used for neural network computing, but GVP is suitable for floating-point operations and is used to process traditional computer vision and millimeter wave radar data, which will play different roles in autonomous driving.
In the CV3 chip, a GPU unit is also integrated to process the vehicle's surround view camera data, which has an hash rate of about 100 GFLOPS.
In addition, the CV3 chip integrates isPs, video decoding, hardware security units, and rich interface management.
In terms of performance, the CV3 chip can achieve an energy efficiency ratio of 10eTOPS/W, which is higher than that of the same level of products in the market. This means that in the same use case, CV3 may only need to actively dissipate heat, but other products need liquid cooling, air cooling and other passive heat dissipation methods.
▲In the future, a number of CV3 chips with different positioning will be launched
Feng Yutao introduced that at present, the CV3 chip has only unveiled a top-level chip, and the evaluation sample of this chip is expected to be provided to developers in the first half of this year. After that, according to the market demand, we will launch mid-range and entry-level products respectively, which can meet the needs of different car companies.
Second, the integration of 4D radar algorithms will focus on the Chinese market
In November last year, Amba completed the acquisition of millimeter-wave radar company Oculii for US$307.5 million (about 1.96 billion yuan), and Has since become a wholly owned subsidiary of Ambarco.
In the CV3 chip, the algorithm of the Auco 4D millimeter wave radar is also integrated.
Xi Jianjun, vice president of marketing in Amba China, pointed out that most of the perception integration solutions on the market are post-integration, and many problems in the post-integration method have been found in the industry. Therefore, CV3 is more suitable for pre-fusion, that is, the fusion of pixels perceived by the camera and the point cloud perceived by radar for the original data set. In this way, the confidence of perception can be improved, and the probability of false detection and missed detection can also be reduced.
▲Jianjun Xi, vice president of marketing in Ambar China
He also mentioned that Auco's 4D millimeter wave radar relies on existing hardware and the algorithm integrated with CV3 to achieve stronger perceptual performance. For example, the effect of an angle radar with low-line lidar after fusing with camera data is very close.
▲Perception picture based on CV3 chip
Xi Jianjun said: "The current 4D millimeter wave radar is a hot product in the market, and many ordinary millimeter wave radars will be upgraded to 4D millimeter wave radar this year, and will play a greater role in the future." ”
In the face of fiercer competition in the autonomous driving industry, Feng Yutao and Xi Jianjun both said that the Chinese market will be one of the most important markets for Ambar.
Feng Yutao said that the iteration speed of China's automotive industry is faster than that of other markets, and passenger cars, commercial vehicles and special vehicles all have AI chip landing scenarios. In the face of the Chinese market, Amba is very optimistic, but also actively responds to the challenges of the Chinese market.
He said that over the years, Ambarco has accumulated a large number of IP, in terms of performance, power consumption, cost performance are excellent, and CV2 series products to achieve mass production on the car, has accumulated a wealth of experience in the car.
In addition, compared to most self-driving chips, CV3 also integrates the algorithm of 4D millimeter wave radar, which has more obvious advantages.
Third, the previous generation of products on the car Rivian vision algorithm is the core competitiveness
When it comes to getting on the car, Ambarco's chip products have been applied to many models of the Great Wall, Dongfeng and Dongfeng Nissan, realizing functions such as driver monitoring DMS and driving recorder.
At the same time, using a number of Amba CV2 chips, Rivian, a new car-making company invested by Amazon, has created an L2-level autonomous driving system driver+ based on CV2 chips, and mass-produced it in an R1T pure electric pickup truck. Motional, a joint venture between Hyundai and Aptiv, also uses domain controllers based on multiple CV2 chips to achieve L4 level autonomous driving. In addition, Arrival, a new car-making force, has also used Amba's CV2 chip to create its L2 level automatic driving system.
▲A number of car companies have built automatic driving systems based on CV2
Based on the experience of mass production of CV2 chips, Aba's CV3 autonomous driving chips will have stronger competitiveness.
Feng Yutao said that visual algorithms will be the core competitiveness of CV3. The reason for this is that the amount of information in visual information is the most dense, because human beings basically rely on vision to complete perception in driving, and they must also be supplemented by hearing, but auditory perception is less.
When building an autonomous driving system, vision is the most important perception system. Therefore, from its inception, Ambarco has focused on the development of visual algorithms.
In the autopilot industry, Tesla attaches great importance to visual algorithms, and even does pure visual perception, cutting off the millimeter-wave radar in mass production cars. In addition, Tesla also sends raw data from camera perception to self-driving computers and data center calculations to achieve more accurate perception.
Ambar is making similar efforts at the same time. Feng Yutao said that in the future, sensor fusion will develop in the direction of raw data pre-fusion, which will also be a major feature of CV3 chips.
In response to the future of the automatic driving computing power dispute, Feng Yutao and Xi Jianjun both said that the competition will become more and more fierce.
Feng Yutao said that the battle of computing power is like an arms race, from the perspective of the development of artificial intelligence, it is difficult to see the end of the struggle for computing power. In the foreseeable future, the car will become a data center on four wheels, and the increase in computing power will depend more on the progress of material science and the progress of chip manufacturing technology.
Xi Jianjun believes that the battle for computing power will become more intense, the car itself will be relatively limited due to the limitations of volume and power consumption, and the importance of the energy efficiency ratio will become higher and higher. At the same time, autonomous data centers will have greater demand for computing power.
Conclusion: The battle for autonomous driving computing power will become more intense
Today, the car has become a data center on four wheels. Especially in the autopilot chip industry, Nvidia, Qualcomm and other chip giants have entered the game, coupled with the rise of domestic chip companies, after the old CV chip companies such as Ambar enter the market, the competition in the autopilot chip industry is becoming more intense.
In the face of competition from autonomous driving chips, Ambarco is using a variety of technologies such as its visual algorithm and self-developed IP core to achieve higher computing power and efficiency. At the same time, the fusion of millimeter wave radar algorithms to achieve pre-perception fusion has multiple advantages at the hardware and algorithm levels.
In today's more fierce struggle for automatic driving computing power, the route taken by Amba CV3 may have a greater advantage.