The development of AI chips is inseparable from the maturity of artificial intelligence technology. Since its birth in 1956, artificial intelligence has experienced three major waves. In the 21st century, due to the improvement of computer performance and the generation of massive data, as well as breakthroughs in machine learning and CNN technology (Convolutional Nerual Networks, Convolutional Neural Networks), algorithms, computing power and data all meet the commercialization requirements of artificial intelligence. , Artificial intelligence has ushered in a stage of rapid development.
In fact, the rapid development of the artificial intelligence industry is inseparable from the only physical foundation at present – the chip. It can be said that “there is no AI without chips”, and whether chips with ultra-high computing power that meet market demands can be developed has become an important factor for the sustainable development of artificial intelligence.
In recent years, the AI chip industry has developed rapidly, and many companies have deployed. However, from the perspective of the start, development and maturity of chips, artificial intelligence chips are still in their infancy.
According to different application scenarios, AI chip design can be divided into three parts: cloud training, cloud inference, and terminal inference. Among them, the cloud training chips are mainly based on NVIDIA’s GPU, and the new competitors are Google’s TPU, as well as Xilinx and Intel, which are deeply involved in FPGA. In terms of cloud inference, representative companies include AMD, Google, NVIDIA, Baidu, Cambrian, etc.
In terms of terminal inference, due to the gradual explosion of demand for mobile terminals, autonomous driving and other application scenarios, the layout companies include traditional chip giants and start-ups, such as Qualcomm, Huawei Hisilicon, Horizon, Cambrian, Bitmain, etc.
It is not difficult to find that in the market structure, although traditional chip giants currently occupy the dominant position in the AI chip market. However, the difficulty of landing AI chips is a common problem that plagues giants and newcomers.
Source: Tsinghua University Future Chip Innovation Center
Huang Chang, co-founder and vice president of Horizon, told Yiou Science and Technology: The reason why AI chips are difficult to implement is that everyone has encountered a common technical bottleneck, which is also the so-called “von Neumann bottleneck”.
One of the keys to improving the performance of AI chips is to support efficient data access. In a traditional von Neumann architecture, data is fetched from memory outside the processing unit and written back to memory after processing. The AI chip itself is based on the von Neumann architecture, and it’s perfectly fine to use simple functions.
However, due to the speed difference between the computing unit and the storage unit, when the computing power reaches a certain level, since the speed of accessing the memory cannot keep up with the speed at which the computing unit consumes data, the additional computing unit cannot be fully utilized, that is, the so-called Von Nuo The Iman bottleneck, or “memory wall” problem, has long plagued computer architecture.
At present, the common method is to use hierarchical storage technologies such as cache (Cache) to alleviate the speed difference between operation and storage as much as possible. However, the amount of data that needs to be stored and processed in an AI chip is far greater than previously common applications. All this makes the von Neumann bottleneck problem more and more serious in AI applications. “It is no exaggeration to say that most of the hardware architecture innovations proposed for AI are fighting against this problem.” Huang Chang added.
However, it is precisely because of the technical difficulties of artificial intelligence chips that both giants and emerging talents are on the same starting line, which provides a good “track” for domestic enterprises to surpass. This also prevents traditional giants from taking advantage of their existing advantages and quickly throwing off their opponents.
On June 20, 19, Cambrian launched the second-generation cloud “Siyuan 270”; on June 21, Huawei released the artificial intelligence mobile phone chip “Kirin 810”; on July 3, Baidu released the artificial intelligence chip far-field voice The interactive chip “Honghu”; on October 29, Horizon released the AIoT edge computing artificial intelligence chip “Rising Sun II”.
It can be found that the layout of domestic enterprises in the field of AI chips has begun to take shape, and there is a fighting force. But if you want to win the world, there are still some deficiencies that need to be improved.
In response to the development of domestic AI chips, Ni Guangnan, an academician of the Chinese Academy of Engineering, has repeatedly stated that the threshold for chip design is extremely high, and only a few companies can afford the cost of research and development of mid-to-high-end chips, which also restricts innovation in the chip field. my country can learn from the successful experience of open source software, lower the threshold for innovation, improve the independent ability of enterprises, and develop domestic open source chips.
“Open source software is becoming the mainstream of the current software industry, and the chip industry can also adopt the open source model.” Ni Guangnan emphasized that in terms of chip development, the new RISC-V instruction set is a new method that can reduce the cost of processor chip IP. model. Enterprises can freely use RISC-V for CPU design, development, and add their own instruction sets for expansion. RISC-V has a good effect on the optimization of the current AI chip architecture and cost control.
Regarding the architecture of AI chips, there are actually many remarkable cases in Chinese enterprises, such as Huawei’s DaVinci architecture, Cambricon-X architecture, Kunyun Technology’s CAISA architecture, Horizon’s Bernoulli architecture, etc. .
Compared with the architecture of artificial intelligence chips, my country should pay more attention to the integrity of the industrial chain of artificial intelligence chips.
The latest equipment and processes for manufacturing chips in my country are many generations behind the international advanced level, so some artificial intelligence chips need to be sent overseas for manufacturing and packaging. This will result in insufficient chip production and high prices. As a result, many downstream products using its modules cannot be mass-produced, resulting in a vicious circle, which is not conducive to the development of the industry.
As a pioneer in the domestic edge-side AI chip field, Canaan Technology has mastered the 16nm process technology as early as 2016. The reason why the current AI chip process technology is still 28nm is mainly due to price and shipment constraints.
CCID Consulting’s “White Paper on the Development of China’s Artificial Intelligence Chip Industry” shows that the scale of China’s artificial intelligence chip market has maintained rapid growth. In the cloud field, the global cloud market accounted for 17.0% in 2018; it is expected to reach 22.15 billion yuan in 2021, with a CAGR of 51.23%. In the terminal field, it will reach 8.41 billion yuan in 2021, with a CAGR of 59.3%.
In the face of such a vast market, I hope that domestic enterprises can concentrate on breaking through the bottleneck and win the world.
Author: Zhang Weichao