Recently, major Internet companies have followed up on ChatGPT:
On February 2, Microsoft announced that all its products will integrate ChatGPT capabilities across the board;
On February 7, Google followed up with the launch of the conversational robot Bard;
On February 7, Baidu officially announced the name of its own ChatGPT-like product as "Wen Xin Yi Yan";
On February 7, Qihoo 360 disclosed that it has used such products internally, and plans to launch a demo version of ChatGPT-like technology as soon as possible.
…
It is not difficult to find that ChatGPT exploded, announcing the arrival of a comprehensive AI era. As computing power is one of the troikas driving the operation of AI large models, the arrival of the comprehensive AI era also means the official opening of an era of computing power.
1. Domestic GPU companies
The underlying chip is the hardware foundation that supports the AI server. An AI server is mainly composed of a CPU and an accelerator chip, where the accelerator chip can be a GPU, FPGA, or ASIC.
In the AI server, each part performs its own duties, among which the CPU is mainly responsible for basic computing tasks such as logical judgment, task scheduling and control; GPU is suitable for general parallel computing and can provide higher computing power for AI training tasks; FPGA has low latency When compared with general integrated circuits, ASICs have lower power consumption and better performance, and can be used to set a specific framework for AI training tasks.
Among the AI chips used in large-scale commercial use today, GPU was the first to be used in the field of AI computing and is the most widely used and mature AI chip in the field of artificial intelligence. GPU is a graphics processing unit. Since it adopts a parallel architecture and more than 80% of it is a computing unit, it has super high computing power and deep learning capabilities. For example, if the sensor is compared to the senses of an adult, then the GPU can be compared to the human brain.
At present, the giants in this field are Nvidia and AMD; domestic listed GPU companies include: Hangjin Technology, Haiguang Information, Jingjiawei, etc.; unlisted companies include Biren Technology, Moore Thread, etc.
2. Domestic FPGA enterprises
Compared with GPU, FPGA has stronger average computing power and lower power consumption. It is suitable for multi-instruction and single-data flow analysis, but not for complex algorithm calculation. In addition, a major feature of FPGA is semi-customization. Users can change the direction of current flow or the gate array through hardware programming to achieve different logic functions.
In recent years, due to the tightening of restrictions on Chinese semiconductors by the United States, the urgent need for domestic FPGA replacement and the rapid iteration of domestic FPGA manufacturers' products have led to the rapid growth of domestic FPGA manufacturers. It is reported that from 2018 to 2021, the compound growth rate of Anlu Technology and Ziguang Guowei both exceeded 100%. In addition, according to Frost & Sullivan's data, the localization rate of domestic FPGA manufacturers is less than 5%, and there is a huge space for localization in the future.
At present, Xilinx and Intel in the United States are the duopoly in this field, while domestic FPGAs are still in their infancy. Cloud Semiconductor, Hejingwei Qili, etc.
3. Domestic ASIC company
ASIC refers to an integrated circuit designed and manufactured in response to specific user requirements and the needs of a specific electronic system. The hardware algorithm cannot be modified after production. Compared with general integrated circuits, it has smaller volume, lower power consumption, improved reliability, and high performance. Improvement, confidentiality enhancement, cost reduction and other advantages.
However, the disadvantage of ASIC is that the initial investment cost is high and the reproducibility is average. Therefore, only when the usage is large enough can the initial investment be allocated and the cost reduced. Previously, with the widespread application of autonomous driving functions, the demand for related ASIC chips rose rapidly.
At present, the representatives of ASIC chip manufacturers that have been listed in China are Cambrian and Montage Technology; among the unicorns, Horizon, Black Sesame Smart, and Suiyuan Technology are all ASIC chip manufacturers, and other ASIC chip manufacturers include Huawei HiSilicon.
