With the vigorous development of cutting-edge technologies such as artificial intelligence, the demand for various computing power has grown rapidly. AI chips have attracted much attention as the core component of computing power.
At present, high-performance computing chips have risen to become a key factor in the competition between the digital economy and artificial intelligence, and even determine the victory or defeat of the global computing power competition.
According to Bloomberg's forecast, the world's top five data center businesses are expected to achieve sustained revenue growth by 2025. At the same time, the global AI chip market also continues to expand, with Gartner predicting that the global AI chip market will reach $67.148 billion by 2024, with a compound annual growth rate of 23.23% from 2022 to 2024.
With the acceleration of domestic substitution, the mainland AI chip industry has ushered in high-speed development opportunities.
Pay attention to [Leqing Industry Observation] and gain insight into the industrial pattern!
Overview of the AI chip industry
Computing power chips are the actual bearers of AI computing and are the key components to improve computing power.
With its high performance, AI chips effectively support the operation of complex algorithms and meet various real-time processing needs, and are widely used in many aspects of artificial intelligence.
From the perspective of technical architecture, AI chips can be divided into GPU, FPGA, and ASIC, according to different network locations, AI chips can be divided into cloud chips and edge chips, and according to application goals, AI chips include training chips and inference chips.
Diagram of the size of the AI chip market:
Source: IDC
As a digital chip, GPU follows the general rule that the smaller the manufacturing process, the better the performance. The GPU industry chain actually constitutes an important part of the advanced process digital chip industry chain. GPU chips have excellent computing power and high efficiency, while maintaining low power consumption, especially good at handling parallel computing tasks in artificial intelligence, machine learning and other applications.
In terms of the FPGA chip market, AMD and Intel dominate the market, and the global market competition pattern is stable. The products launched by domestic manufacturers such as Huawei HiSilicon, Suiyuan Technology and Cambrian are close to the level of similar products of Google in terms of overall performance.
Compared to GPUs, ASIC chips are about 10 times more efficient in training and about 100 times more efficient in inference. With the increasing maturity of large-scale models, the market size of ASIC chips is gradually expanding. Unlike GPUs and FPGAs, the ASIC market has not yet formed an obvious competitive pattern, and domestic manufacturers are increasing their R&D efforts.
AI chip market competition pattern and leading combing
The global multimodal model continues to make breakthroughs, and domestic and foreign technology giants continue to pay attention. Domestic Internet giants such as Baidu and Alibaba continue to increase investment, and various manufacturers are also accelerating the research and development of AI chips to further improve the performance and efficiency of models.
AI chip industry chain map:
Source: iiMedia Consulting
Overseas manufacturers participating in the global competition in the computing chip market include giant technology companies such as NVIDIA, Intel, Qualcomm, AMD, Google and Microsoft.
In the field of GPU and FGPA, foreign manufacturers, especially American companies, occupy a dominant position in the market.
Nvidia has an 84% share of the desktop discrete GPU market, Intel has a 68% share of the global PC GPU market, and Nvidia and AMD have a 19% and 13% market share, respectively.
Nvidia started by developing graphics processors, and in 1999 released the world's first official GPU, the GeForce 256. Since then, the company has been deeply engaged in the field of GPU. In addition, NVIDIA launched a new Blackwell architecture chip platform this year.
NVIDIA has increased AI computing power by 1,000 times in the past 8 years:
In 2023, chip manufacturers such as Qualcomm, MediaTek, and Intel will also launch chips with higher AI performance, providing hardware computing power support for the operation of large models on terminal devices, thereby promoting the application of AI models in terminals.
From the perspective of the domestic AI chip market, domestic products are mainly used for inference, and a few are used for training. The limitation of AI training chips has exacerbated the urgency of China's development in high-process chip design and foundry.
Some representative manufacturers of the domestic GPU industry chain are sorted out:
In terms of domestic computing power, Huawei's Ascend 910 chip is already comparable to NVIDIA's A100 in terms of single-card computing power. Huawei and iFLYTEK have cooperated to build the Ascend 10Ka cluster, in which Feixing-1 is the first large-scale computing platform in China that can train trillions of floating-point parameter large models jointly released by the two parties based on the Ascend ecosystem. #人工智能##算力##财经##科技##服务器##芯片##英伟达#
epilogue
In recent years, the volume of data in the field of artificial intelligence has shown an explosive growth trend, and the number of parameters of algorithm models has increased exponentially. The demand for AI chips in computing centers with accelerated computing as the core is expanding, and it has huge market potential and development space in the long run.
Pay attention to [Leqing Industry Observation] and gain insight into the industrial pattern!