“The future development core of network communication has become cloud service or cloud brain, and Pengcheng cloud brain will be a very important support in the field of network communication in the future.” Gao Wen, academician of Chinese Academy of engineering and director of Pengcheng laboratory, pointed out at the Opening Forum of cite2021.
Academician Gao Wen shared the concept of “cloud brain”, introduced the construction of Pengcheng cloud brain, a major infrastructure, and shared its role in the field of network communication and its promotion to the development of artificial intelligence.
Pengcheng yunnao I
The first generation of Pengcheng cloud brain was completed in 2019. Based on NVIDIA’s chip, it is positioned as a special high-specification computing platform for model training.
It has supported the high-standard AI training tasks of Shenzhen CDC, Shenzhen Traffic Police Bureau and other departments, Huawei, BGI, Yuntian lifeI, Kuangshi and other enterprises, as well as Peking University, Tsinghua University, Shenzhen University, University of Waterloo, Canada, Leibniz Supercomputing Center, Germany and other teams.
Pengcheng yunnao II
The second generation Pengcheng cloud brain is under construction, and its positioning is an international advanced intelligent super computing platform based on domestic chips.
The second generation of Pengcheng yunnao started from the preparation of the project proposal in March 2020 and the trial operation was started on October 15, 2020, including project initiation, feasibility study, preliminary design, optimization, construction, startup and trial operation, which was completed in 9 months.
“This equipment was made with RMB 4.2 billion in nine months, which should not be thought of anywhere,” academician Gao Wen pointed out.
Moreover, Pengcheng yunnao II completely adopts Huawei’s architecture and uses Huawei’s Kunpeng 920 and shengteng 910 machines, of which 4096 are used for 910 and 2048 are used for Kunpeng 920. In fact, it has reached 1024p floating-point computing power, surpassing the international better GPU computing speed in many performance aspects in the same period.
In the second generation Pengcheng cloud brain, every four AI processors shengteng 910 constitute an AI server, so there are 1024 AI servers; Every 256 AI servers form a cluster, so there are four cluster architectures. The four clusters use a fully interconnected high-speed network and adopt a completely unified architecture. It is the strongest machine for artificial intelligence training in the world.
According to reports, Pengcheng yunnao empowers the application from the source innovation, supports the forward-looking technology research and development of leading enterprises, and leads the technology product research and development of small, medium-sized and micro enterprises.
In order to make the ecological environment built by Pengcheng cloud brain better, Pengcheng laboratory has successfully organized a global artificial intelligence competition. Starting from the third competition this year, we hope to run on Pengcheng cloud brain.
“With Supercomputing Center, why do you need Pengcheng yunnao“
The rapidly growing amount of data requires the growth of computing power. Supercomputing is an important aspect of international competition.
“However, supercomputing can not solve all problems. Supercomputing solves scientific computing problems. It requires high precision and speed. In addition, it also needs special support for intelligence. If supercomputing is used as artificial intelligence service, the efficiency may be only one hundredth of that of AI server, that is, ‘big horse pulls small car’, which will cause huge waste,” Gao Wen said. Therefore, cloud brain, a cloud facility built for cloud services, came into being.
Figure: Pengcheng yunnao planning.
“With the continuous growth of data, computing power also needs to keep up. In fact, there is a growing ‘scissors gap’ between data and processing power, and there is an urgent need for an intelligent platform with ultra-high computing density and large-scale computing power,” Gao Wen said. “Pengcheng yunnao aims to provide cloud services to process these data.”
Now, the two technological development routes of Artificial Intelligence Computing and supercomputing are cutting into each other’s direction, and may eventually slowly integrate together. The industry needs to find a good solution to balance the problems encountered in the two directions.
Pengcheng laboratory is using Pengcheng yunnao for this exploration.
Editor in charge: PJ