Fugaku Supercomputer Tops TOP500 Rankings, Showcasing ARM Processor Power

2020‑06‑22 marked a historic day when the ARM‑based Fugaku supercomputer outperformed all competitors in the four TOP500 benchmark tests, overturning common perceptions of ARM processors.

The 55th TOP500 list, released by RIKEN Center for Computational Science (R‑CCS) in Japan, shows Fugaku achieving a Linpack (HPL) performance of 415.5 petaflops.

Fugaku supercomputer at Riken

Fugaku employs 158,976 A64FX SOC processors, each a 48‑core Arm v8‑A design with an additional 4 auxiliary cores, optimized for scientific computation.

The A64FX is the sole processor supporting ARM's Scalable Vector Extensions (SVE), enabling vector units from 128 bits up to 2048 bits, so any software compiled for SVE can automatically scale.

In the HPCG (High‑Performance Conjugate Gradient) benchmark, Fugaku ranked first with 13.4 Pflop/s, while the second‑place Summit achieved 2.9 Pflop/s.

Fugaku outperforms the current runner‑up Summit by a factor of 2.8; its single‑precision and half‑precision performance exceeds 1,000 petaflops (1 exaflop), a range often used for AI and machine‑learning workloads.

Summit, built by IBM, delivers 148.8 petaflops (HPL) with 4,356 nodes, each node housing two 22‑core Power9 CPUs and six NVIDIA Tesla V100 GPUs, interconnected via Mellanox dual‑channel EDR InfiniBand.

Sierra, located at Lawrence Livermore National Laboratory, ranks third with 94.6 petaflops, using a similar architecture of 4,320 nodes, each with two Power9 CPUs and four V100 GPUs, also linked by Mellanox EDR InfiniBand.

China's supercomputers did not reach the top three; the country's total TOP500 performance is 2.23 exaflops, up from 1.65 exaflops six months earlier, largely thanks to Fugaku’s dominance. The new entry threshold for the list is now 1.24 petaflops.

In terms of system count, China still leads with 226 machines, followed by the United States (114) and Japan (30). However, the United States leads in aggregate performance with 644 petaflops versus China’s 565 petaflops.

Accelerators are present in 144 systems (135 of which use NVIDIA GPUs), a figure that has remained stable over the past six months.

Architecturally, x86 remains dominant: 481 of the 500 systems use x86 CPUs (469 Intel, 11 AMD, the remainder Hygon). Only four systems employ ARM processors—three with Fujitsu A64FX and one with Marvell ThunderX2.

Regarding interconnects, Ethernet is used by 263 systems, InfiniBand by 150, and custom networks by the rest. Although Ethernet dominates numerically, InfiniBand‑based systems contribute 803 petaflops of performance compared to 471 petaflops from Ethernet systems.

Related links:

第54届超算TOP500排名，中美表现优异

解析超算TOP500列表，美国5大系统冠领榜单

2017全球超算Top500和Green500榜单出炉

Statement: Thanks to the original author for their hard work. All reproduced articles are credited; please contact us if any copyright issues arise.

Recommended Reading: For more architecture knowledge, refer to the “Architects’ Technical Alliance Bookstore” e‑book collection (32 titles) and click “Read Original Article” for details.

Promotion: Order the “Full Architecture Technical Pack (All)” now to enjoy free updates during the pandemic; price is reduced to 168 CNY (originally 240 CNY).

Warm Reminder: Scan the QR code, follow the public account, and click “Read Original Article” to obtain the “Full Architecture Technical Pack (All)” e‑book details.