On March 9, 2016, AlphaGo defeated Go master Lee Sedol after more than three hours of battle. It was the first time AI showed its amazing power. The Alphago used 1,202 CPUs, 176 GPUs, approximately 8 million parallel cores, and cost $3,000 in electricity. But Lee just needs two steamed buns, a dish of kimchi and a bottle of mineral water it's OK.
It can be seen that AI at that time only waited for the number of piles to be added due to the lack of calculation power, and the energy consumption was very high. In 2017, an upgraded version of Alphago beat international Go master Ke Jie, again attracting attention.
[Baidu Driverless Car]
Baidu founder and Chairman Robin Li Yanhong is seen on the big screen in a driverless car driving on Beijing's Fifth Ring Road when Baidu President Lu Qi introduces the "Apollo" program at Baidu's AI developer conference in Beijing, July 6, 2017. In the footage, Li is seen sitting in the passenger seat, not touching the steering wheel, but being controlled by AI. "The car runs smoothly and feels very good," Mr Li said.
Baidu's self-driving car, Apollo, is said to use AI chips from Intel and Nvidia. Self-driving cars need more than 10 billion kilometers of testing before they hit the road. If AI chips are not used and pure human beings are employed, 100 cars need to be tested 24/7 for 100 years. Apollo can run millions of kilometers a day with an AI chip. ChatGPT's final power consumption is around 3640 Pf-days, and it would take at least 7-8 data centers with a $3 billion investment to support ChatGPT. In terms of model training computing power, at least tens of thousands of Nvidia GPU A100s are needed. Taking GPT-3 with a parameter scale of 175 billion as an example, the cost of one model training exceeds 12 million US dollars.
From these three events, we can see that the computing power demand of AI chips is growing rapidly and approaching the limit of the computing power of existing chips, which forces chip manufacturers to iterate and upgrade quickly. Advanced AI chips represent powerful computing power, which can support the continuous updating and iteration of artificial intelligence and ultimately share a lot of mental and physical labor for human beings.
In general, ChatGPT will stir up the collaborative development of the underlying semiconductor industry, HPC, cloud computing and other industries while blowing the whirlwind of artificial intelligence applications around the world, so as to facilitate the complementary development of related hardware and software and ecology, and even promote the recovery of the semiconductor industry and the sustainable long-term development of related industries.