Artificial intelligence is currently the hottest technology investment project, especially in the United States, where the amount of investment is unimaginable to outsiders. An AI cluster integrating over 100,000 top GPUs has become the norm, and figures like Elon Musk have even proposed an artificial intelligence system integrating 1 million or even millions of top GPUs. What is the limit on the number of GPUs that can be integrated into an AI system in the future? It is estimated that no one can predict.
Is this path the right one? I believe there are issues— the power system cannot support it, and the energy consumption is too high, leading to a disproportionate input-output ratio. One A100 GPU consumes 300 watts; how much power would a system integrating 1 million A100s require? Additionally, there are CPUs and a series of supporting products, whose power consumption is far beyond what an ordinary power system can meet, and the system’s heat dissipation is also a significant problem.
Where does the problem lie? It lies in the chip design architecture.
The so-called GPU is a graphics processing unit, and its design starting point is the computer graphics rendering capability, which has no direct correlation with artificial intelligence. It was merely the accidental system integration by OpenAI that ignited the artificial intelligence business. It is similar to how Pfizer’s Viagra was not originally designed to assist male erection—its initial purpose was to treat cardiovascular diseases, but it inadvertently became a drug to aid male erection.
On October 13, the journal Nature Electronics published the latest research results from the team led by Sun Zhong at Peking University’s Institute of Artificial Intelligence: the new type of artificial intelligence chip designed by this team achieves a computational throughput 1,000 times that of the most advanced AI chips currently available. Once this chip is put into use, it will completely revolutionize the currently designed artificial intelligence systems.
Analyzing from the history of computer development, the research and development of artificial intelligence chips is still in its infancy, similar to the early vacuum tube computers, followed by transistor computers, integrated circuit computers, large-scale integrated circuit computers, and ultra-large-scale integrated circuit computers, etc.
The future path of the artificial intelligence revolution is long, and perhaps it will never have an end. I speculate that after 20-30 years of development, the computational throughput of a future artificial intelligence chip may exceed that of today’s so-called million-card GPU integrated systems. The history of human computer advancement has already proven this.