Currently, a new wave of artificial intelligence has erupted, with massive parameters generating a demand for large computing power. OpenAI estimates that the demand for computing power will double every 3.5 months, nearly 10 times annually. The computing power level of AI big models is significantly in short supply, and high barrier AI chips are significantly benefiting.

With the increasing maturity of artificial intelligence technology and the continuous improvement of digital infrastructure, the commercial application of artificial intelligence will be further implemented, driving the rapid growth of the AI chip market. According to industry research databases, the global artificial intelligence chip market size will reach $72.6 billion by 2025. In the next few years, the market size of artificial intelligence chips in China will maintain an average annual growth rate of 40% to 50%, and by 2024, the market size will reach 78.5 billion yuan.

Overview of AI Chip Industry

According to Xinyu data, compared to high-performance servers and basic servers, AI servers are often more expensive on chipsets (CPU+GPU). The cost of AI server (training) chipsets accounts for up to 83%, while AI server (inference) chipsets account for 50%, which is much higher than the proportion of general server chipsets.

AI chips mainly include graphics processors (GPUs), field programmable gate arrays (FPGAs), specialized integrated circuits (ASICs), neural mimicry chips (NPUs), etc.

The GPU is the core unit of a graphics card, which is a single instruction, multi data processor. Adopting a large number of computing units and ultra long pipelines, it has technological advantages in accelerating the graphics field.

FPGA integrates a large number of basic gate circuits and memory, utilizing gate circuits for direct computation and fast speed. Users can freely define the wiring between these gate circuits and memory, change the execution plan, and adjust to the optimal operating effect. Compared to GPU, it has higher flexibility and lower power consumption.

ASIC is a customized chip designed for specific purposes and user needs, with the advantages of small size, low power consumption, and higher reliability. In the case of large-scale production, it has the characteristic of low cost.

GPU: The Rigidity of Dedicated Computing

GPU (Graphics Processing Unit): It was first proposed by Nvidia when it released the NVIDIAGeForce256 (GeForce256) graphics processing chip in August 1999.

It is a microprocessor specifically designed for image acceleration and general-purpose computing on personal computers, workstations, game consoles, and some mobile devices (such as tablets, smartphones, etc.).

GPU can serve as a training platform for deep learning, with the following advantages: 1. GPU servers can directly accelerate computing services and can also communicate directly with the outside world; 2. GPU servers are used in conjunction with cloud servers, with cloud servers being the main focus, while GPU servers are responsible for providing computing platforms; 3. Object storage COS can provide GPU servers with Big data Cloud storage services.

Its advantages are high performance and usability, while its disadvantages are high power consumption. It has the core barrier of hardware technology and can be used for both GPT model training and inference processes.

The industry chain of the GPU industry mainly involves three links: design, manufacturing, and packaging.

Industry giants are mostly concentrated overseas and occupy the core of various links in the industry chain, playing a decisive role in the global GPU industry.

GPGPU: A highly challenging field

GPGPU servers are currently the most mainstream choice for AI accelerated servers.

The core value of GPU is reflected in graphics and image rendering, and the focus of GPU is on computing power. Although it has evolved from the architecture of GPU, there are obvious differences in the focus.

When designing the GPGPU architecture, the acceleration hardware units designed by the GPU for graphics processing were removed, while the GPU's SIMT architecture and universal computing units were retained, making it more suitable for high-performance parallel computing and able to use higher-level programming languages, which greatly enhances performance, ease of use, and universality.

With the rapid increase in demand for AI accelerated servers, GPGPU shipments are expected to experience significant growth. IDC predicts that the demand for GPGPU chips in China will reach 3.2995 million in 2025, and the shipment volume will achieve significant growth.

GPGPU is a highly challenging field. In the Chinese market, according to IDC data, Nvidia will occupy more than 80% of the market share of China's accelerator cards in 2021, which will lead the market.

Among the domestic enterprises, Haiguang, Huawei Hisense, Cambrian, etc. are all major enterprises in the research and development of domestic coprocessors. The Sugon GPGPU product Shensuan-1 has achieved commercial application in 2021, with sales of 239 million yuan and shipment of 12400 pieces in the first year, becoming one of the first domestic GPGPU chips to achieve commercial sales.

FPGA

FPGA (Field Programmable Gate Array) chips integrate a large number of basic gate circuits and memory, with flexibility between CPU, GPU, and ASIC. Before the hardware is fixed, it allows users to flexibly use software for programming.

In the field of artificial intelligence, FPGA chips can serve as accelerators to accelerate the hardware level computing speed of artificial intelligence algorithms.

FPGA chips are adept at parallelizing operations in the spatial dimension, which is very suitable for the operational needs of neural networks. Therefore, they can significantly improve the computational speed of artificial intelligence algorithms. When facing computationally intensive tasks in the field of artificial intelligence, the characteristics of FPGA chip pipeline parallelism and data parallelism can also greatly improve the operational efficiency of the system.

Overseas manufacturers dominate the global FPGA market, Xilinx and Intel form a Duopoly, and domestic enterprises continue to increase the layout of FPGA chips, with huge growth space.