At present, large models are gradually moving from single-modal to multi-modal, and the demand for GPU-based computing power is exploding rapidly.
The global GPU market is showing a rapid growth trend under oligopoly, with a compound annual growth rate of more than 30%, and the main market share is occupied by American companies such as NVIDIA, and domestic GPU companies have also shown a vigorous development momentum.
At present, GPUs have the highest computing power among hardware, and are the most suitable hardware to support AI training and learning. With the continuous development of AI technology and the growth of application demand, GPU has become a hot "hard currency" in the technology track.
Pay attention to [Leqing Industry Observation] and gain insight into the industrial pattern!
GPU industry overview
At present, the field of AI chips is mainly divided into three categories: GPU, FPGA, and ASIC.
From a market perspective, GPUs dominate the AI chip market due to their superior versatility and flexibility.
GPUs have an absolute advantage in training speed, which is 32 times faster than CPUs for the same throughput training volume.
Since NVIDIA introduced the GeForce 256 chip in 1999 and proposed the concept of GPU, this chip has been widely used in fields that require massively parallel computing, such as machine learning, deep learning, etc.
GPUs contain a large number of arithmetic logic units (ALUs), which are originally designed for graphics computing, but their powerful parallel processing capabilities provide good acceleration for deep learning. GPUs are ideal for computationally intensive tasks that require a large number of matrix operations and convolution operations during neural network training and inference.
CPU vs. GPU:
Source: Architects Technical Alliance
GPUs can be divided into discrete GPUs and integrated GPUs.
Diagram of discrete and integrated GPUs:
资料来源:CGDIRECTOR
GPU is a type of digital chip, and its performance is closely related to the size of the process. Typically a smaller process means that the transistors inside the chip can be more closely aligned, resulting in higher integration, lower power consumption, and improved performance.
The GPU industry chain is essentially an advanced process digital chip industry chain, in order to continuously pursue smaller processes to improve product performance.
GPU market competition pattern and leading combing
The global GPU market presents a clear echelon distribution.
The GPU market is mainly monopolized by three giants: Nvidia, Intel, and AMD.
As the absolute leader in the GPU field, NVIDIA has a global market share of about 87%. At the just-concluded 2024 GTC, Nvidia's Blackwell B200 GPU undoubtedly reflects the forward-looking architecture of the new product, and the powerful performance of the new product leaves a lot of competitors in the dust.
In addition, Intel has a solid market share in the integrated graphics field by virtue of its market share in the CPU market.
The European Union, Japan, and South Korea are in the second echelon, of which the European Union and the United Kingdom have advanced chip equipment manufacturing technology, while Japan and South Korea seize market share by virtue of their advantages in semiconductor material R&D and manufacturing technology.
Overall, the global GPU market presents a three-strong pattern, with Nvidia, Intel, and AMD all having significant technical and market advantages in their respective fields.
Source: Xingxingcha
In the current process of import substitution in the domestic GPU industry chain, leading companies such as Jingjiawei have accelerated domestic substitution in the design process.
In the packaging and testing process, Tongfu Microelectronics, Changdian Technology and Huatian Technology and other companies have successfully undertaken the packaging and testing business of GPUs, demonstrating the domestic capabilities in the advanced process of 14nm and below nodes.
Zhaoxin and Loongson continue to develop GPU products outside the CPU field. In addition, Cambrian, Denglin Technology, Biqian Technology, Muxi and other major general-purpose GPU products.
Some domestic GPU layout manufacturers:
Source: Science and Technology Innovation Board Daily
Huawei's GPU leads the domestic substitution
Huawei's GPUs usually use in-house developed architectures. In terms of the number of GPU cores, Huawei's high-end processors may integrate multiple GPU cores, a design that can significantly improve the parallel processing capabilities of the GPU.
At the same time, Huawei's GPUs also support industry-standard graphics APIs, such as OpenGL ES and Vulkan, which ensures a wide range of compatibility with various applications and games, bringing users a richer and smoother visual experience.
According to reports, Huawei's computing power GPU shipments have reached 100,000 units in 2023, and although the production capacity has increased in 2024, it will still not be able to meet the huge demand of the market. According to public information, the current order demand for Huawei's computing power GPUs has reached millions of pieces. According to multiple agents, since its listing in August 2023, the price of Huawei's computing power GPU has undergone at least two price increases, and the current market price has reached about 120,000 yuan. At present, the domestic computing power chain based on Huawei's Ascend processors is speeding up across the board. #GPU##英伟达##科技##人工智能##财经##算力#
epilogue
With the continuous iterative upgrading of large models in China and the tightening of the supply of high-end computing power market, GPUs are becoming key hardware under the continuous outbreak of AI, which brings huge development opportunities to the domestic AI computing power industry chain.
Pay attention to [Leqing Industry Observation] and gain insight into the industrial pattern!