According to market news, ByteDance is developing an open platform that allows users to create their own chatbots. The company is racing to ride the generative artificial intelligence (AI) wave. According to an internal memo, the "bot development platform" will be launched as a public beta at the end of this month.
The artificial intelligence chatbot program ChatGPT reached 100 million users within two months of its launch, becoming the fastest-growing consumer application in history. It uses generative artificial intelligence technology (AIGC), which is characterized by learning to summarize existing data and then perform deductions and creations to create brand new content. The training and daily operation of ChatGPT are inseparable from the computing power support provided by AI chips.
The concept and market status of AI chips
The AI chip is a module specifically designed to handle a large number of computing tasks in artificial intelligence applications. It has the characteristics of high performance, low power consumption, low latency, and high parallelism.
According to the chip technology architecture, AI chips currently have two development paths: one is to continue the traditional von Neumann computing architecture and accelerate hardware computing capabilities, mainly using GPU (graphics processor), FPGA (field programmable gate array), Three types of chips are represented by ASIC (Application Specific Integrated Circuit); the other is a brain-like chip that subverts the traditional computing architecture and uses brain-like neural structures to improve computing power. It is still in the exploratory stage.
According to the location of the chip in the network, it can be divided into cloud AI chips, edge and terminal AI chips.
Depending on the goals of the chip in practice, it can be a training chip and an inference chip.
With the popularity of AI applications and the increasing demand for computing power, the demand for AI chips is also expanding. Data from research institutions show that China’s AI chip market will be 36.8 billion yuan in 2022 and will reach 142.5 billion yuan by 2027, with a compound annual growth rate of more than 30%.
ChatGPT brings incremental market
The emergence of ChatGPT has driven a large number of technology companies to invest in the research and development of large models. According to public data, the total computing power consumption of ChatGPT-3.5 training reaches as high as 3640PF-days (that is, if one quadrillion calculations are performed per second, 3640 days of calculations are required). Such a huge computing power consumption requires the purchase of at least 6,000 mainstream AI chips. It is conservatively estimated that at least five domestic companies are developing large models with accuracy comparable to ChatGPT, which will bring an incremental AI chip market of at least US$300 million.
All links in the AI chip industry chain in mainland China have unlimited potential
With the development of my country's high-tech industry and the intelligent transformation and upgrading of all walks of life, the application fields of mainland AI chips continue to expand, and its market size shows an accelerating growth trend. However, the development of the mainland AI chip industry is relatively late, and it is not as good as the international advanced level. There are certain differences. Data shows that in 2022, mainland China’s domestic self-sufficiency rate for AI chips will be less than 10%.
The explosion of artificial intelligence applications represented by ChatGPT has provided new opportunities for the development of the mainland's AI chip industry. But Rome was not built in a day. The mainland AI chip industry needs to narrow the gap with overseas countries. There is room for improvement in all links of the industrial chain:
In terms of IC design, the RISC-V architecture has obvious advantages in the field of AI chips. Its open source feature reduces development costs and there is no risk of authorization suspension. In addition, its customizable and scalable features can provide higher operating efficiency, lower energy consumption and more flexible design to better meet the needs of AI computing. Therefore, more AI chip products will adopt the RISC-V processor architecture in the future, which will further promote the development of the RISC-V ecosystem.
In terms of IC manufacturing, under the current international environment, there is great resistance to the production of AI computing chips in overseas Foundry factories. Domestic chips with more than 600G interconnect bandwidth and 600T computing power are gradually tending to be produced in Mainland Foundry factories, which also promotes the gradual development of Foundry factories. Advanced process technology and increased production capacity.
In terms of IC packaging, in the post-Moore era, the chip manufacturing process is close to the physical limit. Vigorously developing advanced packaging processes such as chiplets and interconnecting multiple chips at high speed is one of the ways to improve computing power.
For AI chips, there are two simplest and most effective ways to increase computing power. The first is to increase the computing unit, that is, to increase the chip area. The second is to shorten the physical distance between the memory and the computing unit, increase the bandwidth between the memory and the computing unit, solve the memory bottleneck, reduce the waiting time of the computing unit, and reduce data handling. The key to the success of AI chips is not the chip itself, but the packaging outside the chip.
There is an unwritten rule in the semiconductor industry that the bare die area of a single chip should not exceed 800 square millimeters. If it exceeds 800 square millimeters, the cost will increase rapidly and it will not be practical. To break through the 800 square millimeter limit, the semiconductor industry uses a technology called Chiplet, which can be called a small chip. Multiple small chips are organically connected to form a large chip with commercialization potential in a substantial sense. Therefore, highly integrated chip packaging is the biggest problem facing AI chips.
In this regard, the head of marketing of Yisiwei Computing Technology said: "Promoted by large models such as ChatGPT, AI technology is expected to have a profound impact on all walks of life, indicating that a huge leap in productivity is coming. Various AI applications require low cost, High energy efficiency, flexible design, and ecologically open AI chips are the advantages of the RISC-V architecture. In the future, as the development of AI chips accelerates, the RISC-V architecture will play an important role."