At today’s annual ceremony of “2023 Science and Technology Ranking”,Zheng Weimin, academician of the Chinese Academy of Engineering and professor of Tsinghua University, delivered a speech on the large model training computing power system.In his speech, Zheng Weimin mentioned NVIDIA GPUs and domestic AI chips. He said that NVIDIA hardware has good performance and a good programming ecology. Everyone likes to use it, and many people use it.
But the problem is that they no longer sell to China and cannot buy them anymore. The price has doubled or tripled since December last year.It’s still hard to get a card now.
Regarding domestic AI chips, Academician Zheng Weimin said that Nvidia will no longer sell them to us.Domestic cards must focus on supporting this matter.After making a domestic card, it can be used almost the same as an NVIDIA card, or even slightly worse.
At present, some domestic chips have been made, but users don’t like to use them.The main reason is that the ecosystem of domestic cards is not good.
To change the relatively bad situation of the system, we need to make ten pieces of software: programming framework, program acceleration, communication library, operator library, AI compiler, programming language, scheduler, memory allocation system, fault-tolerant system, and storage system. If you do these well, everyone will like to use them.
As long as domestic AI chips reach 60% of the performance of foreign chips, users will be satisfied if the ecosystem is well established;If the ecology is not done well, even if the hardware performance is 120% of others, no one will like to use it.
Finally, Zheng Weimin concluded that we must vigorously carry out research on large-scale model infrastructure based on domestic systems, change the bad situation of the domestic card ecosystem, do a good job in software and hardware collaboration, and make domestic cards good.