The 10 fastest supercomputers are led by one 2.8x faster than the rest

The newly crowned fastest supercomputer in the world buried the reigning champion by turning in speeds that were 2.8 times faster, according to the current TOP500 list of the world’s fastest.

The ARM-powered Supercomputer Fugaku in Japan scoring a speed of 415.5PFLOP/sec as measured when solving a dense system of linear equations, which is known as the High Performance Linpack (HPL).

Summit, the former number 1, turned in an HPL performance of 148.6PFLOP, which is the same speed it reached last fall when it was ranked first in the semi-annual TOP500 scoring.

When turned against a task of single or further reduced precision testing – a workload used in a artificial intelligence applications – Fugaku’s peak performance reached 1 exaFLOP per second, or 1,000 petaflops.

Summit had been ranked number one on the past three TOP500 lists, and Fugaku was only fully assembled May 22, so had never been in the running before. It is a Fujitsu system with 7,299,072 cores and resides a the RIKEN Center for Computational Science in Kobe, Japan, a research institute with 3,000 scientists spread over seven campuses.

Summit, is an asset of the U.S. Department of Energy’s Oak Ridge National Laboratory with 2,282,544 IBM Power9 cores and 2,090,880 Nvidia Volta GV100 cores. It has a theoretical peak performance of 187.66PFLOP/sec and maintains a power efficiency of 11.324GFLOP/watt.

In addition to Summit, Fugaku also bumped down by one notch the next three fastest supercomputers.

Number 3 is now Sierra, built by IBM for the Lawrence Livermore National Laboratory in Calif. It has 1,572,480 cores supplied by IBM Power9 processors and boosted by Nvidia Volta GV100 accelerators, which add another 1,382,400 cores to the mix. Its performance remains unchanged from six months ago at 94.6PFLOP/sec.

Number 4 is Sunway TaihuLight, installed at China’s National Supercomputing Center in Wuxi, with an HPL performance of 93.0 PFLOP/sec. It’s notable for not using any accelerator chips, relying instead on 40,960 Sunway 26010 processors, each with 260 cores. Its power efficiency is 6.051GFLOP/watt.

At number 5, Tianhe-2A, has an HPL  of 61.4 PFLOP/sec runs on Intel Xeon CPUs and custom-built Matrix-2000 coprocessors with a total of 4,981,760 cores. Deployed at the National Supercomputer Center in Guangzhou, China, its has a hybrid architecture. The name means Milky Way-2A.

Upstart entries

The next two slots are filled by newcomers to the TOP500 top 10.

Number 6 is called HPC5 and like Fugaku is brand new. It is a PowerEdge system build by Dell and installed by the Italian company Eni S.p.A. It it the most powerful system in Europe, achieving 35.5 PFLOP/sec HPL using NVIDIA Tesla V100 as accelerators and Mellanox HDR Infiniband as its network.

Selene, coming in at number 7 is an NVIDIA DGX A100 SuperPOD installed in-house at NVIDIA in the USA. Unsurprisingly it uses the new NVIDIA Ampere A100 processor for acceleration and a Mellanox HDR Infiniband as network to achieve it’s 27.6PFLOP/sec HPL.

The next on the list at number 8 is Frontera which was new last fall, jumping onto the list then at number 5. It’s powered by Intel Xeon Platinum 8280 processors, and is installed at the Texas Advanced Computing Center of the University of Texas, Austin. It’s peak performance is 23.5PFLOP/sec on HPL.

Number 9 is the fourth entry that is new to the top 10 and called Marconi-100, which is installed at the CINECA research center in Italy. It is powered by IBM Power9 processors and NVIDIA V100 GPUs, employing dual-rail Mellanox EDR InfiniBand as the system network. Marconi-100’s performed at 21.6PFLOP/sec HPL.

The final member of the top 10 is Piz Daint with a rating of 21.2PFLOP/sec HPL. It’s a Cray XC50 system installed, and is equipped with Intel Xeon processors and NVIDIA P100 GPUs for a total of 387,872 cores. It resides at the  Swiss National Supercomputing Centre (CSCS) in Lugano, Switzerland.

China dominates numbers and capacity

Looking at just the top 10 fastest, the U.S. claims four slots, China and Italy two each and Japan and Switzerland one each.

But of all 500 supercomputers listed by TOP500, 45.2% are in China and 22.8% are in the U.S., followed by Japan (5.8%), France (3.8%) and Germany (3.2%).

When measuring who has the largest share of overall supercomputing capacity, China again leads with 25.5% followed by the U.S. at 28.7%, Japan (23.7%), France (3.6%), and Germany (3.1%).