author
Zhang Zhidong
The "core" has no far-reaching worries, but will have near-term worries.
Responsible editor 丨cha Youyin
Edit 丨 Chic
With the deep penetration of the concept of "new four modernizations" such as intelligence and electrification, the track of intelligent driving chips has also become popular with the hot new energy vehicle market.
As the name suggests, the intelligent driving chip is the basic hardware used for intelligent driving of vehicles, even if it is a higher level of automatic driving, it is inseparable from the support of intelligent driving chip computing power.
On the one hand, the increasing sentiment of domestic self-developed chips, on the other hand, the exponential growth of market demand, under the dual role, the track of domestic intelligent driving chips has finally emerged a number of excellent players: Huawei, Horizon, Xinchi Technology, Cambrian, Black Sesame...
Although compared with foreign chip giants such as NVIDIA, Mobileye, qualcomm, etc., domestic players have one or another deficiency, but with the bets of capital, the encouragement of policies, and the iterative update of their own chip technology, chip suppliers such as Horizon and Chip Technology have also evolved their own characteristics.
Or the pursuit of higher computing power, or the pursuit of lower power consumption, or from the beginning to do the software toolchain support...
Chip suppliers such as Huawei and Horizon have found a way out of "self-reliance" in their own way under the siege and interception of foreign chip giants. However, when faced with new challenges, can they still maintain their original intention and "kill" the breakthrough as before?
01
How much do you know about smart driving chips?
From a top-down point of view, according to different functions, automotive chips are often divided into computing and control chips (CPU, GPU, FPGA, ASIC, MCU, etc.), power conversion chips, sensing chips, memory chips, communication chips and other types. The intelligent driving chip belongs to the category of calculation and control chips.
Specifically, smart driving chips tend to have two forms of presentation, one is the assIC's standalone chip, and the other is the integrated SoC (System on Chip).
Intelligent driving ASIC refers to a special chip specially designed for specific intelligent driving algorithms. The intelligent driving SoC chip is a chip that integrates modules such as CPU, GPU, and specially designed ASICs to determine the function of the system.
In fact, compared with traditional discrete chips, SoCs tend to have more advantages in terms of performance and power consumption, so major intelligent driving chip manufacturers are more inclined to such a technical route. However, when it comes to the specific architecture inside the chip, it will have its own bias.
It is understood that there are currently 3 kinds of mainstream intelligent driving SoC chip architectures, namely CPU + FPGA, CPU + GPU + ASIC, CPU + ASIC.
Among them, the relatively unfamiliar FPGA, also known as the field programmable logic gate array, is a chip that can adjust the hardware architecture at any time according to a specific algorithm. Therefore, companies such as Waymo, Baidu Appllo, etc., which often iterate intelligent driving algorithms quickly, use cpu + FPGA architecture intelligent driving chips.
Of course, the other two architectures are also being done by players, such as Nvidia's Xavier and Tesla's FSD, which use the architecture of CPU + GPU + ASIC; while Mobieye EyeQ5 and Horizon journey use CPU + ASIC.
Similarly, these three architectures are inseparable from the "ASIC" dedicated design, even if it is an FPGA, its function is similar to that of ASIC, all of which are accelerated designs for specific AI algorithms. And in terms of differentiation, the "ASIC" module of the intelligent driving chip stands properly in the C position.
For example, NVIDIA Xavier's ASIC module is specifically designed according to the "DLA deep learning + PVA vision accelerator", while Tesla is based on the acceleration of "NPU deep neural network".
In fact, the intelligent driving chip adopts a related special design, which is nothing more than to improve computing power, increase data transmission speed, and expand data bandwidth. After all, when people measure a smart driving chip, they often start from these aspects.
But then again, for a smart chip company, its core competitiveness will only be reflected in the performance of a single chip? I'm afraid not.
02
Core competitiveness = threshold
One point that people often overlook is that the reason why a chip can be popular and irreplaceable, in addition to its strong performance, is also related to the ecology it is bound to. In other words, in addition to the strong performance of a single chip, it is also necessary to achieve serialization, matching, and software ecology cooperation in order to form a real threshold.
"Now the threshold for making integrated circuit chips is very low. Theoretically, give me two or three hundred million, I don't have to make a mobile phone chip without a developer, because all the IP can be bought. ”
Hu Weiwu, the father of Loongson, was blunt in a recent interview, directly pointing the spearhead at the "research is better than buying" argument in the semiconductor industry. Subsequently, the entire industry also had a wide discussion around the questions of "self-research boundaries" and "what exactly is the neck of the card".
Of course, the same sharp problem also falls on the body of the smart driving chip. Then you may wish to take the intelligent driving SoC chip mentioned above as an example to see how much gold content its mouth says about "self-research".
First of all, from the system framework of the Black Sesame Intelligent Huashan No. 2 A1000L in the picture above, the ARM 6-core central processing unit (CPU) comes into view directly. There is no doubt that the CPU module is the "available IP core" mentioned by Mr. Hu Weiwu above.
Although Black Sesame has also made efforts in self-developed IP, such as the image signal processor ISP and the deep neural network accelerator NPU shown in the figure above, we do not know what the core modules such as memory LPDDR4 and dual-core vision DSP in the entire system framework are.
Secondly, there is the problem of matching the chip. Intelligent driving chips will not exist alone, and will certainly be connected with smart cockpit chips, intelligent network chips, etc., for the sharing of data, and the series of intelligent functions of the whole vehicle. In this regard, Huawei, Cambrian, and Xinchi all have corresponding cloud chips and networked chips as supporting devices, and Horizon also has the Rising Sun series chips as a cover, so is the single Huashan series losing this threshold?
Finally, it is necessary to mention the support of the chip software ecology. People like to refer to the software package in the semiconductor field as an iceberg below sea level, but the real situation is indeed very different. From the EDA software for chip design, to the simulation software for chip testing, to the software toolchain supporting the development of algorithms, etc., all need the support of the software ecology.
The reason why NVIDIA can achieve "strong and strong" is because there is a low-level CUDA software ecology as a moat. And it has to be mentioned that as far as the intelligent driving chip is concerned, NVIDIA can not only provide supporting hardware, but also provide a full-stack toolchain, including the subsequent development of virtual test suites (software and hardware), physical prototype test suites, etc.
So, how do smart driving chip companies solve this problem? Horizon's tiangong opening, black sesame's mountain and river platform, etc., may be an answer, but together with NVIDIA's "full stack", it is a slight gap.
03
The competitive landscape has changed
Although the intelligent driving chip has better industrial focus and a strong correlation with the "outlet" intelligent driving, history tells us that only with core competitiveness can we truly base ourselves on the industry and be the "tide maker" of the times.
It is not denied that more and more hot money has poured into the intelligent driving chip track, but the degree of burning money in the semiconductor industry is well known, even if it is "hundreds of millions" and "billions" of financing, how long can it burn?
What's more, for now, it seems that only Huawei and Horizon's intelligent driving chips have been mounted on the mass production models. When there is no healthy flow of funds, only rely on the money raised by the outlet to live, how long can chip companies live?
Coupled with the change of the competitive landscape, the repeated appearance of new and old opponents, even car companies such as Zero Run and Geely can't hold back and personally build chips. There are various later forces to catch up, and there are international chip giants such as NVIDIA and Qualcomm.
The so-called "the more competitive, the fairer" argument, I am afraid that it only applies to those who are strong in the industry.
C-DIMENSION
WeChat | Automobile Commune C dimension