●? Let's talk about the familiar mobile phone processor NPU.
Huawei's self-developed artificial intelligence chip was first applied to mobile phone chips, and later applied to servers and autonomous driving systems. Let's talk about the familiar mobile phone processor equipped with artificial intelligence acceleration unit. Huawei's artificial intelligence acceleration unit is called "NPU", and the English full name is "Neural-network? Processing? Units ",translated as neural network processing unit. NPU can accelerate the functions of face recognition, speech recognition and image recognition on mobile phones, thus providing better experience for mobile phone users.
How big is the gap with foreign countries? I mentioned in "Talking about China Autopilot Chip" that the NPU carried by Kirin 970 processor actually comes from Cambrian Cambricon- 1A processor ip, which was carried by Huawei Mate? 10 mobile phone.
Huawei subsequently released Kirin 980 (installed in Huawei Mate? 20 mobile phone) processor's NPU adopts Cambrian Cambricon- 1H processor IP.
Judging from the parameters published on the Internet at present, the NPU computing power of Kirin 8 10 is 0.83TFLOPS (note: 1TFLOPS= one trillion floating-point operations per second), and that of Kirin 820 is 1.33TFLOPS. The manufacturer did not publish the specific calculated values of Kirin 990 and Kirin 985, but came from Zurich Institute of Technology. According to the Benchmark website, Kirin 990 has higher NPU computing power than Kirin 820. According to the test data, it is estimated that its computing power is above 2TFLOPS. The NPU computing power of Kirin 985 will be slightly higher than that of Kirin 820.
Kirin 7 10A is the logo version of Kirin 7 10, which went on sale on 20 18. This chip has completely independent intellectual property rights and will be made by SMIC, so it is a pure "Chinese core". After Huawei Kirin 7 10 enters the automobile circle and successfully lands, it will seize market share with Qualcomm Xiaolong 820A.
With the launch of these new cars equipped with Xiaolong 820A, a wave of high-tech cockpit trends has been brought. Huawei Kirin 7 10A entered the market at this time, just in time for this wave of market. It is believed that many China brand enterprises will consider adopting Huawei's digital cockpit chips of pure China origin in the future.
Although the performance of Kirin 7 10A is different from that of Huawei's latest Kirin chip, and there is no integrated NPU core, it has no problem to meet the performance requirements of multi-screen display and interaction in digital cockpit. Huawei's entry into the automotive field through the digital cockpit is only a test of water, and further layout of the field of self-driving automotive chips is the key.
●? Huawei Autopilot Network Architecture
Ren Zeng, founder and president of Huawei, said in an interview with the media that Huawei will not manufacture complete vehicles, but Huawei will manufacture networking modules and electronic parts in automobiles, which may be the best in the world. Under this circumstance, Huawei will definitely seize the dominant position in the two fields of car networking and autonomous driving.
If Huawei's ADN target architecture becomes an industry standard, it will become an extremely important profit growth point for Huawei in the future 10. Huawei's ambitious goal is to cover all kinds of communication and computing equipment of high-level autonomous driving system and provide a perfect one-stop solution. In general, it is to achieve the effect of "buying communication and computing equipment for self-driving cars and looking for Huawei is right".
Although this paper focuses on chips, Huawei's advantage lies not only in chips, but also in its network communication. Before talking about chips, I want to give a simple example of vehicle-road coordination to illustrate this point.
This simple example can well illustrate how Huawei can empower autonomous driving through network and communication technologies. It can be said that the future self-driving car is by no means an "island", but a mobile tool that connects everything.
●? Huawei's "Core" Strength and Soft Power of Autopilot
Ok, let's briefly talk about the application of Huawei network architecture in vehicle-road coordination. Let's formally talk about Huawei's autopilot platform and chip. Since the autopilot platform is based on Huawei's artificial intelligence chip, we still need to have a simple understanding of Huawei's artificial intelligence chip first.
The application of Ascension 9 10 is biased towards cloud devices that need extremely high computing power, and the application of Ascension 3 10 is biased towards terminal devices that have certain requirements on energy consumption ratio, but there will be some intersection between them. In the following table, I summarized the information of Atlas series products made by Huawei using the above two chips.
The artificial intelligence acceleration chip needs the cooperation of CPU to really run. Some of the Atlas products mentioned above are equipped with Intel's server-class CPU, and the other part is equipped with Huawei's own Kunpeng 920? CPU chip.
This is quite different from other chip manufacturers who only do B2B business and provide development documents after signing the contract. Regardless of the ability of developers, more developers adopt Huawei's equipment development, which will help to build an artificial intelligence ecological chain belonging to Huawei faster and accelerate Huawei's development in the field of artificial intelligence.
Of course, Huawei will also provide targeted software services for different market segments. Due to the limited space, I won't go into details here. Interested friends can go to Huawei Cloud to learn more. While other chip companies are still struggling to find industry breakthroughs, Huawei has provided comprehensive and complete solutions for various industries. Looking at the list of solutions of various industries in official website, I deeply felt the power of Huawei. Perhaps it is Huawei's strength that is the important reason why other companies are struggling to catch up.
●? MDC autopilot platform equipped with Shengteng 3 10.
Ok, after the chip, let's talk about Huawei's autopilot platform. The flagship product of Huawei's latest autopilot platform is MDC600, which can meet the performance requirements of L4 autopilot for domain controllers. Huawei introduced the MDC300 for L3 conditional autopilot.
The single-chip computing power of Shengteng 3 10 is 16TOPS, the power consumption is 8W, and the energy consumption ratio is 2 tops/w; Tesla autopilot? 3.0 The single-chip computing power of FSD chip on the processing unit is 72TOPS, the power consumption is about 36W, and the energy consumption ratio is 2 tops/w; NVIDIA's latest drive? AGX？ Orion platform equipped with Orion chip, the single chip computing power is 200TOPS, power consumption is 45W, and the power consumption computing power is 4.4 tops/w ... In comparison, the energy consumption ratio of Shengteng 3 10 has caught up with the international mainstream level. At present, Shengteng 3 10 is manufactured by TSMC 12nm process. With the upgrading of production technology to 7nm or even 5nm in the future, there is still room for further improvement in its energy consumption ratio.
At present, MDC autopilot hardware platform and HarmonyOS system kernel of vehicle operating system have successively obtained the requirements of ISO26262 international standard for road vehicle functional safety. ASIL-D grade (hereinafter referred to as "ISO26262? ASIL-D "), followed by Huawei's adaptation? AUTOSAR architecture, development tool chain and autopilot algorithm can also pass ISO26262? ASIL-D, what about Huawei? The MDC autopilot system will be the first in the world to meet the requirements of ISO26262? ASIL-D commercial autopilot system.
Conform to ISO26262? What does ASIL-D mean? This means that 1 time random hardware failure will only occur in the MDC autopilot system on the vehicle every 1 100 million hours, and the safety factor is quite high. The probability that ordinary users encounter random hardware failure when using MDC autopilot system is extremely low.
The system architecture of MDC autopilot platform is extensible. By increasing or decreasing the number of CPU cores, artificial intelligence accelerated cores and IO interfaces, different use scenarios of high, medium and low-end passenger cars from driver assistance to high-end intelligent driving can be met.
Today, Huawei's MDC autopilot platform has signed more than 18 customers, among which SAIC, Geely, Jianghuai, FAW Hongqi, Dongfeng Motor, Suzhou Jinlong, Neolithic and Shandong Haorui Intelligent are all on the list.
From the cloud to the terminal, Huawei has a complete set of autonomous driving solutions, so can it be said that Huawei is invincible? Don't! What Huawei lacks in the field of autonomous driving is massive data of actual road conditions. Without these data to train the autopilot system, the system can only stay in the laboratory and close the test drive. To this end, Huawei also actively cooperates with related enterprises, such as NavInfo, which has map resources, and a number of China vehicle companies to make up for their shortcomings in road big data and further build their own technical barriers in the field of autonomous driving.
●? Editor's summary:
There is no doubt that Huawei's influence in the industry is in the forefront of the world from technology to products. In recent years, Huawei has been sanctioned by the United States. If the status quo continues for a long time, it will have a great impact on Huawei artificial intelligence chips, which this paper focuses on. How to solve the problem is a difficult problem for Huawei. Judging from some current information, Huawei's measures such as building a chip factory in the UK, developing RISC-V architecture processors, and cultivating the HMS cloud service ecosystem are the key to its breakthrough. We will continue to pay attention to the further development of the situation. (Photo/Text/car home? Chang Qinglin)