What opportunities can the phenomenal popularity of ChatGPT bring?

Practitioners in the field of AI believe that ChatGPT is a new starting point for the application of AIGC (AI Generated Content, the use of artificial intelligence technology to generate content), with the continuous improvement of deep learning models, the promotion of open source models, and the possibility of commercialization of large models, AIGC is expected to enter the application explosion period. What is the role of the commercial landing of ChatGPT technology in promoting the semiconductor industry?

Currently, the hottest application of artificial intelligence (AI) is ChatGPT, a chatbot program released by OpenAI. ChatGPT can react to the context when talking to the user, and also supports text output tasks such as writing code, email, script, and copywriting.
Due to ChatGPT’s excellent performance in the Q&A session, the program attracted the attention of the global market as soon as it was launched, and the number of registered users exceeded 100 million two months after its launch. In this context, the huge market potential behind ChatGPT has yet to be tapped, and this technology is expected to leverage a large number of application scenarios.
Practitioners in the field of AI believe that ChatGPT is a new starting point for the application of AIGC (AI Generated Content, the use of artificial intelligence technology to generate content), with the continuous improvement of deep learning models, the promotion of open source models, and the possibility of commercialization of large models, AIGC is expected to enter the application explosion period. What is the role of the commercial landing of ChatGPT technology in promoting the semiconductor industry?
Training AI models requires massive computing resources
Essentially, ChatGPT is a class of AI language models. According to official information, ChatGPT is fine-tuned on the basis of the GPT 3.5 model. ChatGPT and InstructGPT are sister models, both of which use a new training paradigm in the field of Large Language Model (LLM) generation – RLHF (Reinforcement Learning from Human Feedback), that is, to optimize language models based on human feedback in a reinforcement learning way, and the two have only slight differences in data collection settings.
According to information revealed to users by ChatGPT, RLHF is a way of using reinforcement learning to directly optimize language models with human feedback. In fact, it involves multiple models and different training stages, and its techniques can be roughly broken down into the following three points: first, pre-training a language model (LM); Second, aggregate question answering data and train a Reward Model (RM); Third, fine-tune LM with reinforcement learning (RL). In summary, RLHF enhances human regulation of model outputs, and also enables more comprehensible sequencing of results.

Looking at the GPT-3 model before GPT 3.5, the number of parameters is as high as 175 billion, and the amount of data required for training is 45TB. By using the RLHF training method, even if InstructGPT has only 1.3 billion parameters, its output is still better than GPT-3. According to the data released by OpenAI, InstructGPT and ChatGPT are sister models, so it can be guessed that the number of parameters of the two may not be much different.
Perhaps some readers lack a specific concept of the amount of parameters, this article explains by giving a popular example – in September 2020, Microsoft obtained the exclusive license of OpenAI GPT-3, which built a supercomputing center for training GPT-3, and the supercomputer in the center is loaded with 10,000 NVIDIA GPUs. Training GPT-3 consumes 355 GPU-years of Microsoft’s computing power (1 GPU runs for 355 years), and the cost of a single training is as high as $4.6 million. However, as of the date of writing, the author has not queried the information of ChatGPT model training costs.
ChatGPT is an AI model, which is inseparable from the support of computing power – in the training stage, a large amount of corpus data is required to train the model, and in the application stage, a large computing power server is required to support the operation of the model. Even if the number of parameters in ChatGPT can be reduced to billions, training and running it requires a lot of computing resources.
Promote the further development of intelligent computers
The phenomenal popularity of OpenAI ChatGPT has pushed technology companies to accelerate the deployment of ChatGPT-like products. Recently, Google, Microsoft, Baidu and other companies have announced that they will provide AI model services, of which Google will launch a conversational artificial intelligence service Bard supported by the LaMDA model, Microsoft embedded OpenAI’s GPT-3.5 in the search engine Bing, Baidu will promote the new model project “Wen Xin Yiyan”, the initial version will be embedded in the search service.
As more technology companies deploy ChatGPT services, huge computing power will be required to train AI models, and this demand has also made some enterprises see the business opportunities of AI computing power services. On February 10, 2023, Inspur launched AI computing power service products. The company said that based on the computing power infrastructure of China’s leading intelligent computing center, it will provide AI computing resources and supporting services for Chinese customers, support the whole business process of AI model construction, training and inference, and empower generative AI industry innovation.
Of course, deep-pocketed tech giants may build their own computing centers. For example, Microsoft’s supercomputer, the aforementioned device, is used to train hyperscale AI models on the Azure public cloud. It is understood that the center is loaded with more than 285,000 CPUs and 10,000 GPUs, of which the network connection capacity of a single GPU server is 400Gb/s, and the peak computing power of the device can perform 23.5-61.4 trillion floating point operations per second.
Purchasing AI computing power services and building computing centers require the support of large-scale computer equipment.

In further discussion, we must first clarify a concept – computing power represents the strength of the ability to process digital information, and there are great differences between different types of computing power. For example, the computing power unit of a supercomputer is FLOPS (floating point computing power per second), while the computing power unit of an intelligent computer is OPS (operations per second), which are two different concepts.
In addition, the accuracy of computing power should also be considered to measure the level of computing power. The Lookout think tank pointed out that the current Linpack test used to measure supercomputing in the industry tests the “double-precision floating-point arithmetic ability” of supercomputers, that is, the calculation of 64-bit floating-point numbers (FP64). In addition, in the digital precision expressed in binary, there are single precision (32 bits, FP32), half precision (16 bits, FP16), and integer types (such as INT8, INT4). The higher the number of digits, the higher the accuracy, the higher the complexity of the operation, and the wider the application scenarios it can adapt.
Intelligent computer is a special computing power device, which performs well in intelligent computing such as inference or training, but most intelligent computers do not have high-precision numerical computing capabilities; Supercomputer is a kind of general computing power equipment, its design goal is to provide complete, complex computing power, in the high-precision computing power is stronger, the application range is wider, mainly used by scientific researchers for planetary simulation, new material development, genetic analysis and other scientific computing and big data processing.
AI model training only uses intelligent computers, but the current intelligent computer technology is not mature, and it is only applied in pattern recognition, knowledge processing and development intelligence. Although this type of equipment has not yet reached the expected goal, it has made some progress in the fields of text, speech, graphic image recognition and understanding, and machine translation, and related primary products have also been launched.
Drive the development of more AI market segments
As mentioned earlier, the emergence of AIGC is expected to leverage the large-scale landing of AI. At the same time, the layout of related enterprises in multiple AI industry chains will be deeper. For example, the hardware layer includes chips and sensors, of which AI chips mainly include general-purpose GPUs, customizable FPGAs, dedicated ASICs, and brain-like chips. NVIDIA’s Orin chip is based on the general-purpose GPU, Horizon’s Journey 5 chip, which is both an ASIC chip and a DSA (Domain Specific Architecture) chip; Brain-like chips include IBM’s TrueNorth, Intel’s Loihi, Qualcomm’s Zeroth, Xijing Technology’s DeepSouth, Hangzhou Dianzi University’s Darwin, aiCTX’s DynapCNN processor, etc.
Overall, AI chip suppliers mainly include NVIDIA, AMD, Intel, VelociHOST, Jingjiawei, Horizon, Cambrian, Bitmain, Fudan Micro, Xilinx, Altera (Intel), Heterogeneous Intelligence, Google, etc.; The sensor part has suppliers such as Sunny Optics and Hesai Technology; The algorithm layer includes SenseTime, Megvii, Cloudwalk Technology, YITU Technology, Fourth Paradigm, etc.; The application layer includes Hikvision, iFLYTEK, Ruiming Technology, etc.
Although intelligent computers do not pursue too large computing power frequencies, they have high requirements for power consumption and heat dissipation of chips, so low-power FPGA and ASIC chips have greater use in intelligent computers. At the same time, AI chips will also shift from training scenarios for manufacturers to reasoning scenarios for consumers. AI experts in the industry commented that the coordination and unification of high parallel computing capabilities and high versatility of GPUs may be difficult to sustain in the consumer era, and in the future, ASIC chips and Chinese-made GPGPU chips may be able to cut into the MaaS (travel as a service) industry ecology.
Will the development of the server industry chain benefit?
Some readers may think that in theory, more computing power means more computer equipment, and more core devices are needed to build these devices. Does this mean that enterprises’ pursuit of ChatGPT-like technology plays a positive role in promoting the development of the server industry chain?
The author lists some global server industry chain information in Table 2, mainly involving key components and complete machine suppliers.

Server motherboard vendor
Server motherboards are specially developed to meet server applications, requiring high stability, high performance, and high compatibility. The server motherboard suppliers listed in this table are concentrated in China and the United States, such as Intel and Supermicro in the United States; Chinese mainland’s Lenovo, etc., as well as Asus, Gigabyte, MSI, Tai’an (owned by Shenda) in Taiwan, etc.
Server CPU vendor

As of Q1 2023, about 90% of the world’s server CPUs use x86 architecture, and the remaining 10% or so use non-x86 architecture. At present, Intel occupies more than 90% of the market share of x86 server CPUs, and AMD, which is also an x86 architecture camp, although it has been chasing in the field of PC CPUs in recent years, its share in server CPUs is difficult to shake Intel’s position. IBM’s CPU uses the Power architecture, and its global market share is also lower than Intel’s. In addition, Taiwan’s Cyrix (acquired by VIA Electronics), Chinese mainland’s HiSilicon, the Institute of Computing of the Chinese Academy of Sciences, Tianjin Feiteng, Shenwei Technology, etc. also have server CPU products, but the market share of these companies is far from Intel.
· GPU vendor
AI models require a lot of deep learning and machine training, such as GPT-3 with 175 billion parameters, and need to use a large number of NVIDIA V100 (32GB), A100 and H100 (80G) GPU products. At present, NVIDIA A100, H100 GPU products have been adopted by Amazon AWS, Microsoft Azure, Google Cloud, Oracle and other companies. According to IDC statistics, in the field of GPU servers in China, NVIDIA’s market share is as high as 95%, and almost all cloud service providers and supercomputers use NVIDIA’s chips to support AI computing.
In addition, AMD, Intel, VelociHOST and other US-funded companies also produce GPU products; China’s GPU suppliers include Jingjiawei and Changsha Shaoguang (a subsidiary of Hangjing Technology).
Storage related suppliers
The storage-related vendors listed in this table are involved in memory, external memory (hard disks) and memory interface chips.
At present, the memory interface chip has been upgraded to the DDR5 generation, and the suppliers are Montage Technology, Rambus, and IDT. Although DDR5 with a rate of 4800MT/s penetrates faster in PC laptops than servers, only when the transmission rate reaches 6400MT/s, the PC needs to be equipped with DDR5 memory interface chips. Therefore, the current DDR5 memory interface chip is more used in servers, and it is expected that during the three-year period from 2022 to 2024, the penetration rate of server-side DDR5 will be 15%, 40%, and 60%, respectively.
Memory suppliers include SK Hynix, Samsung Electronics, etc. in South Korea, Micron Technology and Kingston in the United States, Kingtech and Citon (Taiwan) in China; External storage suppliers include Western Digital (including SanDisk), Seagate, etc. in the United States, Lenovo and Kingtech in China.
Server supplier
In the server part, the author mainly lists the suppliers of China and the United States.
The United States has Dell, HPE, IBM, Cisco, etc.; China has Huawei, Xinhua3, Inspur, Lenovo, Dawning, Shandong Transcendence and so on.
According to the current market judgment, enterprises’ pursuit of ChatGPT-like technology will directly promote the development of the AI chip industry, and the server and its industrial chain as a computing power infrastructure device may have better development opportunities. From March 29 to 30, 2023, AspenCore will hold the International Integrated Circuit Exhibition and Symposium (IIC Shanghai 2023) in Shanghai, where AI vision chip supplier AiXin Yuanzhi, storage-related suppliers Longsys, Dongxin Semiconductor, ISSI, domestic high-performance server CPU supplier Feiteng and other enterprises will participate. At the same time, the EDA/IP and IC Design Forum will also invite well-known EDA/IP, IC design companies at home and abroad to participate in the speech, welcome interested friends click here to register, and come to the venue to communicate.
The benefits for the server industry are not obvious in the short term
Supercomputers and smart computers are made up of multiple servers. Objectively, the pursuit of ChatGPT-like services by technology companies is conducive to the development of the global server industry chain, but this role may be difficult to reflect in the short term. According to analysts’ forecasts, the performance of the global server market in 2023 is not very good.
At the end of January 2023, Jibang Consulting revised its forecast data for the global server market in 2023, and the growth rate was lowered to 1.87%. The agency said that affected by the continued weakness of the global economy, the four major cloud service providers in North America revised down their server purchases in 2023, and the figures may continue to be revised downward, from the largest to the smallest for Meta, Microsoft, Google, and Amazon Web Services. The server purchases of these four companies increased by 6.9% year-on-year to 4.4%, which will affect the annual growth rate of global server shipments to 1.87% in 2023.
In fact, the enterprises that have determined to deploy ChatGPT are mainly some large-scale and powerful giants, after all, the deployment of such services requires great computing power, and the cost of training and verifying models is also very expensive. It is believed that with the further industrialization of more such technologies, different business segments will appear in the market.