ChatGPT has a dual impact on computing power

ChatGPT has been a hot topic in recent months. Behind its continued popularity is the rising demand for computing power for automatic generation of artificial intelligence content. On the one hand, ChatGPT has driven the growth in demand for computing power. On the other hand, ChatGPT has put forward higher requirements for computing power support.

Data shows that the total computing power consumption required for a ChatGPT model training is 3640PF-days, costing about 12 million U.S. dollars, and also requires considerable overhead during operation. . According to the Guosheng Securities report, based on the average number of unique visitors to ChatGPT in January of 13 million, the corresponding chip demand is more than 30,000 NVIDIA A100 GPUs, the approximate computing power cost is 800 million US dollars, and the daily electricity cost is 50,000 US dollars. Around USD. ChatGPT has received a US$13 billion investment from Microsoft, which is the confidence support behind its high computing power investment. It is precisely with the full support of the technology giant Microsoft from funds, future markets, to computing infrastructure (Azure cloud) that OpenAI can complete such a huge project as an independent entrepreneurial enterprise. Computing power requirements are closely related to parameter magnitude. Parameter magnitude is also an important reference for currently measuring the quality of large model training. The larger the number of parameters, the smarter the model, and the greater the overhead that comes with it. This forms the characteristic of high threshold for GPT type. The previous version of GPT-3 was open source, making it easier for other entrepreneurs to carry out new research and development work based on it. Starting from GPT-4, OpenAI open API charges based on byte traffic. Whether it is application development or secondary model development based on this, there is a high threshold.

The more interesting thing is that Meta’s LLaMA source code was leaked. As a result, the so-called "Llama Family" ("Llama" can be translated as "Alpaca") large model emerged. Developers who want to obtain open source resources can get a new option besides GPT-3.

The "large model"'s high dependence on computing power has at least two major impacts.

First, this is no longer a game of "grassroots" entrepreneurship, nor is it a game of "Copy To China" for Internet entrepreneurship in the past, but a competition of core capabilities of "real swords and real guns", from capital, technology to Comprehensive assessment of corporate strategic determination. In the traditional industrial era, China started from the processing industry and gradually moved up to the research and development of high-value core technologies. However, in fact, it is still a certain distance from the most advanced level in Europe and the United States. Especially in categories such as materials science that require long-term technical accumulation, the gap is still large.

Entering the digital era, the consensus among Chinese industry circles has become to start directly with core technology research and development.

It has become a trend for the United States and China to become less dependent on each other in science and technology. Although it is only the chip industry that is being constrained now, if "big models" develop into a factor of production, we will definitely encounter the same problem. Therefore, China must have its own "big model" core technology. Seeing this inevitability, Chinese technology companies have also become a trend to invest heavily in “big models”.

If small and medium-sized technology companies do not have the ability to develop native "big models", they can use the open source models of GPT-3 or LLaMA for development or secondary development to form their own "big models". Application development companies rarely develop native "big models". Unless they have extremely high confidence in the sales market of their applications, they will succumb to huge computing power cost pressures and instead pay API usage fees to develop related applications using ready-made models. .

Secondly, the construction of “Counting in the East and Counting in the West” comes at the right time and may usher in big business opportunities.

The western region is rich in "green power" resources. Under the overall plan of "Eastern Digital and Western Computing", the western hub is mainly responsible for backup and storage; the eastern region is responsible for real-time computing in smart cities, industrial Internet and Internet services. The work will be settled at the hub in the eastern region. The "big model" has huge requirements for computing power. Eastern IDC occupies computing power resources on a large scale, which has three major disadvantages: high cost (high electricity bills in the east), high emissions (the proportion of green electricity in the east is small), and squeezing out computing power for applications with low latency requirements. resources, which creates new business opportunities for computing power centers in the west.

On the one hand, placing the training calculations of "large models" in IDCs in the western region has higher requirements for improving the performance of local communication networks. The throughput of massive data is a problem that needs to be solved, but the cost of solving this problem is far less than the difference in IDC usage fees.

On the other hand, technology company giants have all put forward their own ESG (environmental social responsibility) goals, and carbon emissions are the most important indicator. The operation of "large models" and large computing power will produce a large amount of carbon emissions, and the demand for "green computing power" will become more urgent.

The above is the detailed content of ChatGPT has a dual impact on computing power. For more information, please follow other related articles on the PHP Chinese website!