TOP500 Supercomputer Rankings 2018: Performance, Architecture, and Global Trends

The 52nd TOP500 list, published in November 2018, places five U.S. Department of Energy (DOE) supercomputers among the top ten, with the summit and Sierra systems from Oak Ridge National Laboratory (ORNL) and Lawrence Livermore National Laboratory (LLNL) occupying the top two positions.

Since its debut in June 2018, Summit has increased its High Performance Linpack (HPL) score from 122.3 to 143.5 Petaflops, while Sierra improved from 71.6 to 94.6 Petaflops, moving from third to second place. Both systems use IBM‑built hardware with Power9 CPUs and NVIDIA V100 GPUs.

Sierra pushes China’s Sunway TaihuLight, which held the TOP500 lead for two years, down to third place with an HPL performance of 93.0 Petaflops; TaihuLight was developed by China’s National Research Center of Parallel Computer Engineering and Technology (NRCPC).

China’s Tianhe‑2A (Milky Way‑2A) ranks fourth with a Linpack score of 61.4 Petaflops, featuring a proprietary Matrix‑2000 chip that replaces Intel Xeon Phi accelerators.

The fifth‑ranked system is Piz Daint, a Cray XC50 at the Swiss National Supercomputing Centre (CSCS), delivering 21.2 Petaflops using Intel Xeon CPUs and NVIDIA Tesla P100 GPUs.

Trinity, a Cray XC40 operated by Los Alamos and Sandia National Laboratories, reaches 20.2 Petaflops and ranks sixth, being the only top‑10 system that employs Intel Xeon Phi processors.

Japan’s AIST AI Bridging Cloud Infrastructure (ABCI) system, powered by Fujitsu, ranks seventh with 19.9 Petaflops, using Intel Xeon Gold CPUs and NVIDIA Tesla V100 GPUs.

Lenovo’s SuperMUC‑NG, located at Germany’s Leibniz Supercomputing Centre, holds the eighth position with 19.5 Petaflops generated by over 311,000 Intel Xeon cores.

USA’s Titan, a Cray XK7 at ORNL, falls to ninth place with 17.6 Petaflops, leveraging NVIDIA K20x GPU accelerators.

IBM’s BlueGene/Q system Sequoia at LLNL ranks tenth with 17.2 Petaflops.

China now accounts for 227 systems (45 % of the total), while the United States has dropped to 109 systems (22 %). Although the U.S. has fewer machines, its average system performance remains higher (38 % of total performance vs. China’s 31 %).

Japan ranks third in system count with 31 machines, followed by the United Kingdom (21), France (18), Germany (17), and Ireland (12); other nations have fewer than ten each.

The top ten manufacturers by system count are Lenovo (140), Inspur (84), Sugon (57), Cray (49), HPE (46), Bull (22), Fujitsu (15), Huawei (14), Dell EMC (13), and IBM (12).

Intel remains the dominant CPU supplier, holding 95.2 % of the processor market in TOP500 systems, with only three systems using AMD CPUs.

ARM‑based supercomputers entered the list for the first time: HPE’s Astra, built on 125,328 Cavium Thunder X2 cores, achieved 1.5 Petaflops and placed 205th.

Overall, 137 systems now use accelerators or co‑processors, up from 110 six months earlier; NVIDIA continues to lead with 64 systems using P100 GPUs, 46 using V100 GPUs, and 12 using Kepler GPUs.

Ethernet is the most common interconnect (252 systems), followed by InfiniBand (135), custom interconnects (64), and Intel Omni‑Path (43). The aggregate TOP500 performance grew from 1.21 Exaflops to 1.42 Exaflops.

The Green500 list, which ranks systems by energy efficiency, places Japan’s Shoubu system B (RIKEN ZettaScaler‑2.2) first with 17.6 Gigaflops/W, followed by NVIDIA’s DGX SaturnV Volta (15.1 Gigaflops/W) and ORNL’s Summit (14.7 Gigaflops/W). Other high‑efficiency systems include AIST’s ABCI, TSUBAME 3.0, Sierra, and several European machines.

In the HPCG benchmark, which complements Linpack by measuring memory and network performance, Summit and Sierra again lead, achieving 2.93 and 1.80 HPCG‑Petaflops respectively. Fujitsu’s K computer ranks third (0.60 HPCG‑Petaflops), Trinity fourth (0.55), and ABCI fifth (0.51). Additional top‑ten HPCG performers include Piz Daint, Sunway TaihuLight, Nurion, Oakforest‑PACS, and Cori.