On November 12, 2021, in Beijing, the third China Super Computing Conference (ChinaSC 2021) was successfully held. Two major lists were released at the meeting: China’s high-performance computer performance TOP100 list and the international artificial intelligence performance ranking AIPerf500.
(1) Top 100 list of high-performance computer performance in China
The 2021 China High Performance Computer (HPC) performance TOP100 list is jointly released by the High Performance Computing Professional Committee of the Chinese Computer Society, the High Performance Computing and Mathematical Software Professional Committee of the China Industrial and Applied Mathematics Society, and the China Intelligent Computing Industry Alliance.
2021 is the twentieth time to release the “Top 100 List of High-Performance Computers in China”.
Zhang Yunquan, a researcher at the Institute of Computing Technology of the Chinese Academy of Sciences and one of the publishers of this list, said that since the list was first released in 2002, it has experienced international giants dominate the list, China’s supercomputing is self-reliant, and gradually achieves autonomy and controllability. Witnessed the development of China’s supercomputing business.
The average performance of all 100 supercomputing systems on the list in 2021 will increase by 79% compared to 2020. The performance of the 2021 “1st place” is 1.34 times that of 2020, and the performance of the “100th place” in 2021 is 1.06 times that of 2020.
Zhang Linbo is a researcher at the Institute of Mathematics and Systems Science, Chinese Academy of Sciences, and director of the State Key Laboratory of Computing for Science and Engineering.
The test performance that ranked first was 125PFLOPS, and the system peaked at 240PFLOPS.
The second-place test performance is 93 PFLOPS, and the system peaks at 125 PFLOPS.
The third-place test performance was 87 PFLOPS, and the system peaked at 160 PFLOPS.
The fourth-place test performance is 61 PFLOPS, and the system peaks at 100 PFLOPS.
On the list, there are not only “old faces”, but also “new faces”. “Sunway Taihu Light” and “Tianhe No. 2A”, which were the top two on the list in 2020, are old faces. Interestingly, on this list, 8 of the top 10 are new faces.
The new host systems are as follows: A total of 29 host systems will be submitted in 2021, of which Lenovo will submit 18 host systems, Beilong Chaoyun/Dell Chaoyun will submit 3 host systems, and Beilong Chaoyun/Intel Chaoyun will submit 1 host system Mainframe systems, the server provider submits 7 mainframe systems.
In terms of performance distribution, there are 7 host systems with performance greater than 10 PFLOPS.
There are 10 host systems with performance in the range of 3~10 PFLOPS.
There are 30 host systems with performance in the 2.5~3 PFLOP range.
There are 51 host systems with performance in the range of 2~2.5 PFLOPS.
There are 2 host systems with performance in the range of 1.5~2 PFLOPS.
Among them, there are a large number of host systems with performance in the range of 2-3 PFLOPS. One of the reasons is that the price is affordable for manufacturers.
The sixth-ranked console system has attracted much attention, but it has to sell. This host system is not to be underestimated, the highlight is efficiency first, it has strong double-precision floating-point computing performance.
Note: One TFLOPS (teraFLOPS) is equal to one trillion (=10^12) floating point operations per second.
One PFLOPS (petaFLOPS) is equal to one quadrillion (=10^15) floating point operations per second.
The sixth-place host system is equipped with 992 pieces of Sunway’s new generation processor SW26010Pro, and uses Sunway’s latest network architecture, Sunway Network, for high-speed interconnection. It is one of the outstanding representatives of the new generation of Sunway computers. The test performance is 12.59 PFLOPS, the system peak is 13.91 PFLOPS, and the Linpack efficiency is as high as 90.3%.
The supercomputing system at the bottom of the list this time is the Inspur TS10000 cluster deployed in a laboratory in Shenzhen. The test performance is 1986TFLOPS, which is 1.06 times that of the 100th place in 2020.
The top three manufacturers share a total of 80% of the system shares on the list, namely, Lenovo 40 systems, Inspur 28 systems, and Sugon 12 systems, with the top three performance accounting for 28%.
In terms of performance share in the application field, in 2021, the High Performance Computing Professional Committee of the China Computer Federation will add a new field of “computing service”, which refers to the application part that uses computing power as a resource to provide computing services to users.
“Hashrate service” has the largest share of system performance, accounting for 46%, followed by supercomputing centers with 24%. Artificial intelligence accounts for 9%, cloud computing accounts for 5%, and short videos account for 4%. These categories are still hotspots in current supercomputing applications.
Regarding the “computing service”, Zhang Yunquan, a researcher at the Institute of Computing Technology of the Chinese Academy of Sciences, commented: “The computing service has suddenly emerged, and China has entered the era of computing power economy.”
This time, it is the 20th time that China’s HPC TOP100 list has been released, which means that this list has gone through 20 years.
(2) AIPerf500, the International Artificial Intelligence Performance Ranking
This time, it was jointly initiated by Zhang Yunquan, a researcher at the Institute of Computing Technology of the Chinese Academy of Sciences, Professor Chen Wenguang of Tsinghua University, Pavan Balaji, an applied research scientist and technical director of Facebook, and Professor Torsten Hoefler of the Zurich Laboratory, Switzerland. Based on AIPerf (scale artificial intelligence computing power benchmarking program) The international artificial intelligence performance ranking AIPerf500 is the second release.
The developer of the AIPERF evaluation program, Zhai Jidong, a tenured associate professor at Tsinghua University, said: In order to better promote the development of the intelligent supercomputing center and reflect the overall computing power of the intelligent supercomputing center, it is very important to design a scalable intelligent evaluation program. The AIPerf evaluation program is committed to reflecting typical artificial intelligence application scenarios, and at the same time, it has better scalability than similar international evaluation programs, which is very important for evaluating ultra-large-scale intelligent supercomputing centers.
As can be seen from the AIPerf500, the 2021 International Artificial Intelligence Performance Ranking, with the addition of more large-scale clusters, the threshold for the list has increased rapidly. The tenth place in the 2020 list uses only 8 Nvidia V100 (32G) GPUs, while the tenth place in the 2021 list uses 128 GPUs of the same model, and its score is 20.6 times the score of the tenth place last year).
It is worth noting that 15 new AI host systems were added, 13 of which entered the top 20.
As Chen Runsheng, an academician of the Chinese Academy of Sciences, said in his speech: “As we all know, now the entire human race has entered the era of big data, and big data has deeply penetrated into all aspects of people’s production and life. It plays a very important role. For example, in the field of biology that I am familiar with, it has led to the emergence of precision medicine, which will make our medication more accurate, the diagnosis more scientific, and the treatment more effective. Not only that, big data will certainly Profound changes in the paradigm of biomedicine have made us transition from the current focus on diagnosis and treatment to the era of precision medicine and focus on health assurance. Therefore, driven by big data, society is undergoing profound and paradigm shifts in various fields. changes, and big data will inevitably bring big information, it also requires large computing power to mine new knowledge brought by big data. Therefore, in the current situation, computing power is adapted to our entire The needs of human social science development have also adapted to the needs of our country’s economic development and innovation.”