The story begins with edge calculation.
In recent years, Internet companies in China have enjoyed a high reputation all over the world, especially in the field of artificial intelligence, and they are at the forefront of the world. The reason is that because of China's demographic dividend and relaxed environment, users use the mobile Internet to generate a large amount of data, which are analyzed in the cloud, thus establishing a large number of data models, which provide great convenience for the development of various companies. At the same time, these industrial scenes and scales have also cultivated many artificial intelligence talents.
But in the 5G era, the subject of data generation will no longer be just people. With the rapid increase of network speed and the sharp decrease of network delay, the Internet of Things will be widely used. The networked objects, that is, all kinds of intelligent devices, will generate more and more complex data through their own sensors, such as cars, cameras, temperature sensors, drones and so on. The analysis, calculation and modeling of these data will produce greater value. Usually, these data are processed in cloud computing. But with the popularity of artificial intelligence, edge computing (edge? The concept of computing is also increasingly recognized.
What is edge computing? Generally speaking, it is to put the computing work on the device itself that generates data, or on the server that is close to the device, rather than far away from the cloud. Compared with cloud computing, edge computing has several advantages. First, the delay is small. Clouds and terminals usually have network delays ranging from tens of milliseconds to hundreds of milliseconds. For applications with high delay requirements such as autonomous driving and industrial applications. The deployment of artificial intelligence in the cloud cannot meet its demand for delay; Second, data privacy. Some applications do not want to transfer data to the cloud. On the one hand, they are worried that cloud data will be seen by cloud operators, on the other hand, they are worried that they will be hijacked by hackers during data transmission. Finally, reliability. If artificial intelligence is deployed in the cloud, once the network is disconnected, the terminal's artificial intelligence program will not work, which is difficult to meet the requirements for automotive applications that require high reliability, but if artificial intelligence is deployed at the edge, there will be no such problem.
There are many levels of edge artificial intelligence computing. One is artificial intelligence computing on terminal devices (such as mobile phones). This scheme puts artificial intelligence calculation directly on the terminal equipment, which can achieve the lowest delay. However, due to the limited battery capacity of terminal equipment or low tolerance to heat dissipation, artificial intelligence computing on terminal equipment puts forward extremely high requirements for the energy efficiency ratio of AI chips. The other is to assume an edge server near the device. Driverless driving is the best example, because it needs sensors to be fused on unmanned vehicles to do a lot of calculations, which is equivalent to deploying an edge server on unmanned vehicles.
China's first automobile clearance edge computing chip
But it is not easy to set up a server in the car. First of all, you have to make a chip that can still be unique in the field of chips that many giants are involved in. At the same time, it must meet the stability requirements of the vehicle regulations, and finally form a set of feasible solution servers.
Looking at the current global artificial intelligence chip market, there are chip giants such as Intel and NVIDIA, Internet software giants such as Google, and traditional tier 1 suppliers such as Freescale, Renesas and Texas Instruments. The first three companies are big companies in the Internet field, which have invested a lot of money and resources and have great technical advantages in the field of artificial intelligence, while the latter has high reliability, low cost and high technical content. To hoard goods in such competition, Horizon must have its own "killer".
The answer given by the horizon is journey. Its BPU (brain? Processing? Unit) processor is based on ASIC (application specific integrated circuit), which refers to the integrated circuit designed and manufactured according to the requirements of specific users and the needs of specific electronic systems. ? ) is introduced by deep neural network learning optimization, which has advantages in vertical power consumption, cost and computing power.
The research and development route of BPU processor solution is divided into three stages: Gaussian architecture, Bernoulli architecture and Bayesian architecture. Based on Gaussian architecture, Horizon released two chips, of which Journey 1.0 chip can support L2 driver assistance function, and its technical parameters are between Mobileye? Eyes? Q3？ And then what? The power consumption of EyeQ4 is obviously lower than that of the two products, and the typical power consumption is1.5w..
At the 20 18 CES, the Journey 2.0 chip released by Horizon is an advanced computing chip based on Bernoulli architecture, which is oriented to L3 and above autopilot. Based on this architecture, Horizon Technology also launched an advanced autopilot computing platform, Matrix? 1.0, its core chip is the FPGA with the architecture of Journey 2.0, and other auxiliary chips are also integrated on this board.
Matrix? 1.0 supports multi-sensor fusion, and can process four channels of 720P video at the same time, with a real-time rate of 30 frames per second and power consumption controlled at 31w. Matrix 1.0 also lays a good foundation for verifying, developing and further optimizing the Journey 2.0 chip.
At CES in 2020, Horizon exhibited the Journey II chip (Journey? 2) and its matrix? 2 platform. The processor was successfully streamed at the beginning of 20 19, using TSMC 28 nm? HPC+ technology, based on BPU(Brain? Processing? Unit)? 2.0 architecture, integrating two Cortex-A53 cores, providing 4TOPS(Tera? Operation? Per？ The abbreviation of Second, 1TOPS means that the processor can perform one trillion calculations per second), and its performance is more than 10 times that of the same level GPU.
In addition to the powerful computing power under the premise of low power consumption, Zhengtu 2 processor can also provide high-precision and low-latency perceptual output. The accuracy of typical target recognition is over 99%, and the delay is less than 100 milliseconds. In terms of identifying objects, Journey II has been able to identify more than 60 types of targets, and the number of single-frame targets has exceeded 2,000. In addition, the processor can efficiently and flexibly handle various AI tasks, detect and accurately identify various targets in real time, and can fully meet the requirements of intelligent driving scenes such as automatic driving visual perception, visual mapping and positioning, visual ADAS, and the functional requirements of intelligent human-computer interaction such as voice recognition, eye tracking and gesture recognition. Crucially, Zhengtu 2 processor was designed in strict accordance with the reliability requirements of vehicle regulations at the beginning of research and development, and passed the certification of automotive electronic reliability standard AEC-Q 100, becoming the first AI chip of vehicle regulations in China.
Based on the Journey 2 processor, Horizon also launched a new generation of Matrix? 2.0 The self-driving computing platform can improve the computing power by up to 16 times, while the power consumption is only 2/3 of the original. At the same time, it can support video input up to 8 million pixels, and the pedestrian detection distance is as high as 100 meters. It also meets the perceptual computing needs of self-driving fleets and unmanned low-speed cars in different countries and different scenarios. It is expected to be listed this year.
At the same time, at this year's CES, Horizon executives also introduced BPU? Journey 3, the third generation Journey processor based on 3.0 architecture, is specially designed for autopilot and domain controller, and also conforms to AEC-Q 100 and ISO? 26262 automobile classification standard, and Matrix is expected to be released next year? 3.0 Autopilot computing platform will also adopt the third generation processor of Zhengtu, and its computing power will be improved to 192TOPS, supporting ASIL? D's system's ability to apply scenarios.
Two-digit front car fixed-point and Geneva landing debut
In 2020? At CES, vice president of Horizon and vice president of automobile product line told Bao Auto Studio that the Journey II chip development kit is completely ready, which can support customers to design products directly. The chip has obtained the pre-loading fixed points of customers in five countries in the directions of advanced autonomous driving, ADAS and multi-mode interaction, and obtained the pre-loading fixed points of double digits. At the upcoming Geneva Motor Show, domestic brands equipped with Zhengtu 2 chips will be unveiled soon.
Although Zhang Yufeng did not mention the specific brand and model. However, it is not difficult to guess the specific vehicles landing at the Geneva Motor Show from Horizon's partners. According to the materials mastered by Baoche Studio, Horizon's autonomous driving solution directly serves commercial fleets, OEMs and Tier? 1。 Open partners include Bosch, Audi and Changan. Only Changan Automobile and Aichi Automobile are the brands that China participated in this time. At this auto show, the main style of Chang 'an is Uni-T (Chinese name: Gravity), which focuses on the city of design and technology.
As can be seen from the official pictures released by Changan Automobile, the car is equipped with L2-level solutions including millimeter-wave radar and front camera. And the supporting chip behind it is likely to come from the horizon. Zhang Yufeng said that the mass production of ADAS function is a great opportunity for the domestic automobile industry in the future, because the loading rate of this function in China is below 10% at present. For Horizon, the market share growth of ADAS is a huge opportunity. Therefore, the top priority of Horizon's work this year is the mass production of the second generation chips.
But this is not the whole of Horizon's efforts in the field of autonomous driving. According to Zhang Yufeng, Horizon adopts the strategy of "walking on two legs". On the one hand, it cooperates closely with the main engine factory and Tier 1, on the other hand, it also cooperates extensively with commercial motorcades that promote Robotaxi business. When Horizon cooperated with the self-driving taxi fleet, it provided three horizon matrix computing platforms for each car to form a 12 camera system, which was recognized by the market. In terms of afterloading, Horizon has reached cooperation with many well-known travel service providers and operators at home and abroad, including the first car about the car, SK Telecom, etc., and has achieved batch deployment. In addition, Horizon has authorized nearly 1000 L4-class self-driving vehicles for domestic and foreign self-driving manufacturers and Robotaxi operating fleets. This time, I continued to come to the United States to participate in CES in order to further expand its Robotexi support business in North America.
Horizon's mass production of automotive chips in the field of autonomous driving is undoubtedly good news for China's automobile industry as a whole. It can be predicted that the demand for computing power in the automotive and artificial intelligence industries in the future is amazing. In the past few years, we have seen that every level of intelligent driving has increased, and the computing power has almost increased by an order of magnitude. If fully automatic driving is to be realized, the vehicle needs thousands of top computing capabilities. However, at present, the average computing power of products in the automobile market has not reached the top of single digits. Therefore, facing the huge market demand, the strength of China enterprises is very important. If we don't pay attention to this strategic highland, we may still be subject to overseas restrictions like aircraft engines and automobile powertrains. The emergence of China's national power not only means that our strategic highland is expected to be adhered to, but also may attack the city in the world market. Some experts in the industry predict that the journey chip is expected to reach the pre-installation capacity of one million in two years and the goal of ten million in five years.
This article comes from car home, the author of the car manufacturer, and does not represent car home's position.