What new computing power demands does the big model bring?

idcuu 2023-11-29 阅读:380 评论:0
On November 24th, the China Unicom Yangtze River Delta (Wuhu) Intelligent Computing Center project with an investment of...

On November 24th, the China Unicom Yangtze River Delta (Wuhu) Intelligent Computing Center project with an investment of 6 billion yuan began construction;

On November 22nd, the Alte (Wuxi) Intelligent Computing Center for automotive research and development was officially launched;

On November 21st, the China Mobile Intelligent Computing Center (Wuhan) with the first planned 1000PFLOPS intelligent computing capability was launched;

On November 20th, the first domestically produced intelligent computing center project was officially signed in Gui'an;

In addition, land planning related to intelligent computing has been released in places such as Ordos and Nanchang... In the past week, news of intelligent computing related projects has been circulating almost every day. And this is almost the norm in the past year, as the construction of intelligent computing centers has entered an explosive period.

The popularity of intelligent computing is driven by the rise of AI big models and their new demand for computing power.

How terrifying is the computational power consumption of large models?

With the rapid development of artificial intelligence technology, AI big models have become the main driving force for the growth of computing power demand in the world today. These models not only make breakthroughs in theoretical research, but also play an increasingly important role in practical applications. The increasingly popular AI big models are driving the transformation of computing infrastructure worldwide and in China.

We can get a glimpse of how much computing power is required for AI large models from the training computing power consumption of ChatGPT. The OpenAI report shows that training a GPT-3 model with 174.6 billion parameters requires approximately 3640PFlops day of computing power. Although the details of ChatGPT 4 have not been disclosed, according to ARK Invest, the maximum number of GPT-4 parameters is estimated to be 150 billion, and the highest computing power demand can reach 31271 PFlops day.

This is just a training requirement for ChatGPT - after ChatGPT attracted widespread attention, countries around the world are entering the field of large models, and various types of large model products are emerging one after another. In less than a year, China alone has released no less than 200 large-scale model products.

It can be imagined how terrifying the computing power required for these large models will be.

The challenges faced by computing power infrastructure

The surge in demand for computing power has brought challenges to computing infrastructure in two directions.

On the one hand, it is the challenge of computing power production. With the limit of Moore's Law at the level of physical laws, relying on single chip computing power is no longer sufficient to meet the demand for ultra large scale computing power. Therefore, more chips are needed to work together. As a result, the scale of data centers is increasing, and the construction of cluster computing power centers has become the norm.

On the one hand, it is the challenge of the environment required for producing computing power. In the context of the ultra large-scale development of computing power centers, the sufficient and stable energy supply of computing power centers, as well as how to effectively dissipate the heat brought about by massive energy consumption, have also become a practical topic that needs to be faced.

And these are only the computing power required for AI large model training. With the implementation and popularization of large model products, the demand for inference based computing power will enter a growth period, and its increment will exceed the training demand. In the long run, its total computing power demand may be equal to or even exceed the training demand.