Has most computers on list, but trails U.S. in processing power
SAN JOSE, Calif. — The U.S. retook a lead in supercomputer performance on the latest Top 500 list, but for how long remains to be seen. China not only extended its lead in the number of high-performance systems but it is racing to retake a performance lead with an early exaflop-class computer.
The latest results give the U.S. bragging rights but show that China continues to grow in its technological sophistication. It comes at a time of trade tensions between the countries in part due to friction over semiconductor policies.
The list also shows that accelerators, mainly Nvidia GPUs, continue to drive many of the most muscular systems in the world. However, Intel and Pezy Computing, a fast-rising accelerator maker based in Japan, also made significant showings.
The U.S. took the first and third spots in the latest Top 500 after not having an entry above fourth place for a year. The Summit system at Oak Ridge National Laboratory was measured at 122.3 petaflops on the Linpack benchmark, edging ahead of the previous leader, China’s Sunway TaihuLight, at 93 PFlops.
The Sierra system at Lawrence Livermore National Laboratory came in third at 71.6 PFlops. Both Summit and Sierra use IBM Power 9 processors and Nvidia Tesla V100 accelerators in more than 4,300 nodes riding Mellanox Infiniband EDR interconnects.
The U.S. also retook a lead with 38.2% of the overall Linpack performance of systems on the list compared to China at 29.1%. Six months ago, the roles were reversed.
The results are even more pronounced using a new high-performance conjugate gradient (HPCG) benchmark introduced on the latest list. The Summit and Sierra systems lead the HPCG rankings at 2.926 and 1.79 PFlops, respectively.
By contrast, China’s TaihuLight came in sixth place with just 0.481 PFlops. China’s second-most-powerful system, the Tianhe-2, only reached #49 with 0.038 PFlops. All other China systems ranked below #100 using the benchmark.
Linpack has long been seen as a relatively crude measure of raw theoretical performance. By contrast, HPCG uses code that measures seven kinds of operations “designed to stress data movement, which is seen in many real applications,” said Jack Dongarra, a University of Tennessee professor and one of four organizers of the Top 500 list.
Despite losing bragging rights in performance, China retained its lead in the most supercomputers on the list with 206 systems, up from 202 six months ago. The U.S. fell to a historic low with 124 systems, down from 145 in the November 2017 list.
In addition, Lenovo now has the most systems on the list (119), the first time that a China vendor has claimed that role. Hewlett Packard Enterprise, the previous leader with 122 systems, now has 79. It is followed by Inspur, Cray, and Sugon with 68, 56, and 55, respectively.
Although IBM came in sixth with just 19 systems on the list, its presence in the Summit and Sierra systems gave it bragging rights as having the greatest share of overall Linpack performance on the list at 19.9%. Cray and Lenovo followed at 16.5% and 12%, respectively.
China racing toward exaflop goal
While China slumped on the latest Top 500 list, it has been gearing up since 2016 for its next big leap, said Dongarra. The country is widely expected to deliver the first system capable of an exaflop on the Linpack benchmark, perhaps a year before the U.S.
Three teams in China are currently competing to build relatively small prototype systems of about 512 nodes, each using different architectures. Interim results of their work may determine which team will get funding to build a finished exaflop supercomputer, perhaps as early as 2020.
The three projects are led by China server vendor Sugon and the teams behind the Tianhe and Sunway supercomputers, said Dongarra. Universities, research institutes, and national supercomputing centers act as partners.
The Sugon system aims to connect more than 10,000 nodes on a 6D torus network, each node using x86 processors developed in conjunction with AMD. The Tianhe team is expected to use a version of the homegrown Matrix 2000 used in Tianhe-2A, perhaps augmented with Arm-based processors or accelerators. The Sunway prototype may roughly follow the many-core-based approach of TaihuLight.
A year ago, the U.S. government awarded $258 million to an effort that aims to deliver an exascale system by 2021. China and the U.S. share a goal to deliver an exaflop system with a 30-GFlop/W energy efficiency and greater than 500 Gbits/s in interconnect throughput, said Dongarra.
An increasing share of the world’s most muscular systems use accelerators — 110 of the current Top 500 systems, up from 101 six months ago. Nvidia’s GPUs are by far the most popular co-processors, appearing in 96 of those systems, including five of the top 10 systems.
Seven systems use Intel’s Xeon Phi co-processors, and four in Japan use accelerators from Pezy, a relative newcomer.
Interestingly, the Pezy accelerators power the top three systems in the related Green 500 list that measures supercomputers with the highest energy efficiency. The rest of the top 10 on the Green 500 list use Nvidia GPUs as accelerators.
“The Pezy systems are cooled by liquid — that is, the whole system is immersed in a liquid [typically a version of 3M’s Fluorinert], which is much more efficient than air or water through plates,” said Dongarra.
The vast majority of the Top 500 systems use Intel Xeon CPUs as their main processors. Nearly half (247 systems) now use 10-Gbit/s Ethernet as their interconnect, up from 228 six months ago.
The relatively more expensive Infiniband declined from use as an interconnect in 163 systems six months ago to 139 on the current list. Intel’s proprietary Omni-Path interconnect is in 38 systems on the latest list, slightly up from 35 six months ago.
— Rick Merritt, Silicon Valley Bureau Chief, EE Times