How does ChatGPT affect China's chip industry?

In 2023 after the epidemic dissipates, the hottest concept in the technology field is none other than ChatGPT. Bill Gates calls it the second revolutionary technology, Huang Renxun calls it the iPhone moment of AI, and the venture capital community even more With Wang Huiwen, Li Kaifu, Li Mu, Jia Yangqing and other leading figures stepping down one after another, starting a new entrepreneurial journey in the field of large models, this track is already lively enough.

The industry and investment community’s enthusiasm for ChatGPT is based on ChatGPT’s potential to empower thousands of industries.

In terms of UGC, BuzzFeed CEO Jonah Peretti said in a memo to employees that content created by AI will transition from the research and development stage to part of the core business this year, while in other areas On the user production platform, there are already many accounts that use AI personalized production to produce content and have achieved profitability.

ChatGPT can also empower education. On the one hand, it uses AI to create children's picture books, which can be used as an educational extension of fables. On the other hand, it can also help adults do conversational exercises in spoken English, further improving human life and work. The efficiency...

Semiconductors, as the cornerstone industry that provides computing power support for ChatGPT, will naturally be affected by this wave, especially the Chinese semiconductor industry that is clamped down by the United States. How does ChatGPT affect the development pattern of China's semiconductors?

At CITE 2023, which just ended last week, CoreList and CITE jointly held a seminar on AIGC and computing power chips. Some people believe that AI is the main driving force of the next semiconductor cycle. Some people believe that AI is the main driving force of the next semiconductor cycle. It is believed that ChatGPT will trigger the rapid development of the four major subdivisions of computing chips, memory chips, packaging and testing, and PCB. Others believe that the integration of storage and computing will play a huge role in the future.

AI drives the recovery of the semiconductor industry cycle and stimulates the four major areas of the industrial chain

Since the birth of Moore’s Law, the consumer electronics industry has been acting as a driving force The main driving force for the development of the semiconductor industry. After the growth of the mobile phone business slowed down, PCs stimulated a wave of domestic demand during the epidemic and became the main driving force for the continued growth of the semiconductor industry. Now that we have entered the post-epidemic era, the demand for consumer electronics is no longer strong, the semiconductor industry has entered a downward cycle, and new application stimulation is urgently needed to become the driving force.

Yao Ningbo, managing director of Kailian Capital, introduced at the conference that in the past, the semiconductor cycle was driven by mobile phones. Now mobile phone sales have dropped by 10% every year. During the epidemic, PCs have become the main driving force. In the future, new energy vehicles will Become the new driving force.

"AI, HPC and cloud services continue to increase in volume, and computing will become the main driving force in the next semiconductor cycle," Yao Ningbo said.

In addition, Yao Ningbo also emphasized that Samsung’s recent announcement of production cuts means that the semiconductor cycle is about to bottom out, and the next semiconductor cycle will arrive in Q3 or Q4.

Xinbang CEO Ban Keke believes that the opportunities that ChatGPT brings to the chip industry are reflected in the industrial chain, mainly concentrated in four major areas, computing chips, memory chips, packaging and testing industry and PCB industry chain.

As ChatGPT’s attention continues to rise, relevant organizations predict that the global growth rate of GPU will rise from 32.82% to 39%. The surge in demand for computing chips will trigger business changes in wafer factories, from TSMC’s business It can be seen from the revenue that its high-performance computing continues to grow at a high rate, accounting for 41% of the total revenue, surpassing the smartphone business for the first time. The annual growth of Nvidia's data center HPC chip performance will be as high as 200%-250%.

In the field of storage chips, Ban Keke said that HBM, as a new type of CPU and GPU memory chips, has increased in volume and price. It can not only meet ChatGPT’s new demand for storage, but also provide a solution for the gloomy storage market. The industry brings a touch of brightness. In June 2022, SK Hynix announced the mass production of HBM3 and shipped it to NVIDIA, which will be paired with H100 GPU to achieve accelerated computing. Currently, HBM orders from SK Hynix and Samsung have surged, and HBM has become a new competition point in the memory chip industry.

In the field of packaging and testing, the demand for Chiplets, which has become popular in recent years, is even greater.

In the PCB board industry, PCB boards are required for core parts such as motherboards, power backplanes, hard drive backplanes, and network cards in servers. The computing power revolution has increased the use of PCB boards, and it has also promoted the use of PCB boards. Develop towards high frequency and high speed. For domestic manufacturers, Shenghong Technology, the PCB supplier of Nvidia's A100/H100 graphics cards, will directly benefit.

Perhaps in the face of the US ban and the inability to purchase A100, China can only make a castrated version of ChatGPT based on A800. But in addition to GPU, domestic semiconductor manufacturers can also find new opportunities in other links of the industry chain.

The integration of storage and calculation for lane changing and overtaking injects new power into China’s AI large computing power chips

The most essential demand of ChatGPT is the demand for computing power. Although we currently cannot match NVIDIA in the GPU field, we can improve computing power efficiency in another way. Storage and computing integrated technology dedicated to breaking the memory wall has now become a major direction that the industry is looking forward to and that capital is chasing.

Xu Fang, senior vice president of Yizhu Technology, explained at the seminar that from the perspective of the overall environment, the current demand for computing power is increasing, Moore’s Law is approaching the end, and the market’s requirements for energy efficiency ratio continue to increase, but The energy efficiency ratio and computing power density are close to the ceiling. This is the reason and necessity for industrial development to integrate storage and computing.

On the other hand, from the perspective of the law of scientific development, Amdahl's law has always been the basic law of hardware acceleration design. It represents the ability of the processor to improve efficiency after parallel computing. According to the derivation of the formula, it is shown that storage and calculation are integrated The AI ​​acceleration performance is much better than GPGPU. In other words, to a certain extent, the integration of storage and computing can contribute to the improvement of computing power and even have advantages over GPUs.

In addition to companies like Yizhu Technology, which specializes in the development of integrated storage and computing chips, other domestic memory chip companies have also been involved in the integration of storage and computing.

At the CITE conference, memory chip manufacturer Yixin Technology told Leifeng.com(Public account: Leifeng.com) that the training of ChatGPT requires a large-scale data set, and the data first needs to be stored After massive data storage, there is a need for a bidirectional interface with sufficient bandwidth to read large amounts of data, which brings opportunities to the memory chip industry.

"Yixin itself is for storage. We support large-bandwidth data transmission, and the offline end supports data encryption. We can independently complete data storage, calculation, transmission and encryption work. When implemented in specific scenarios, we can help users Strengthen the requirements for data application in scenarios. "

"According to the customer's AI application needs, personalized applications can be trained in a targeted manner, but this puts forward new requirements for storage and computing, so traditional Static storage does not have much value, but it is necessary to extract the stored data and optimize the model to better meet customer needs." Xiong Mingxia, senior director of marketing at Yixin Technology, explained.

This article is reproduced from Leifeng.com. If you need to reprint, please go to the official website of Leifeng.com to apply for authorization.

The above is the detailed content of How does ChatGPT affect China's chip industry?. For more information, please follow other related articles on the PHP Chinese website!