Homepage Rankings and Research Companies Channelcast Marketing Matters CRNtv Events WOTC Cisco Partner Summit Digital 2020 NetApp Digital Newsroom HPE Zone The Business Continuity Center Enterprise Tech Provider Masergy Zenith Partner Program Newsroom HP Reinvent Digital Newsroom Hitachi Vantara Digital Newsroom IBM Newsroom Juniper Newsroom Intel Partner Connect 2021 Avaya Newsroom Experiences That Matter The IoT Integrator Intel Tech Provider Zone NetApp Data Fabric WatchGuard Digital Newsroom

Top 500 Supercomputers: New No. 1 Uses Arm-Based Fujitsu CPUs

British chip designer Arm achieves a major milestone with a new system in Japan becoming the world's fastest supercomputer as Intel and AMD continue to duke it out over the x86 server space. 'If it's an inviting performance-per-dollar formula there, then they eventually get some attention,' an Intel and AMD partner says.

While Intel continues to dominate the world's top supercomputers and AMD's EPYC processors continue to gain ground, Fujitsu's new Arm-based server processors have catapulted over the dueling chipmakers to power a new Japanese supercomputer that is debuting at the top of the prestigious list.

The Fugaku supercomputer, operated by the RIKEN Center for Computational Science in Japan, was named to the No. 1 slot in an updated list of the world's top 500 supercomputers, which was published Monday by Top500 for the opening of the virtual ISC 2020 high-performance computing conference.

[Related: Nvidia's Ian Buck: A100 GPU Will 'Future-Proof' Data Centers For AI]

Top500 said Fugaku is the first supercomputer in the No. 1 slot to be powered by Arm-based CPUs — a major milestone for British chip designer Arm, which is looking to make a bigger splash in the data center with its Neoverse CPU architecture that is being licensed to Amazon Web Services and other partners.

Fugaku, which is powered by Fujitsu's 48-core Arm-based A64FX system-on-chip, consists of nearly 7.3 processor cores and reached 415.5 petaflops of performance in the High Performance Linpack benchmark, outperforming the world's previous fastest supercomputer, Summit, by 2.8 times. A petaflop amounts to one quadrillion floating-point calculations per second.

When running at peak performance for single or further reduced precision math, Fugaku can perform more than 1,000 petaflops, or 1 exaflop — the milestone three U.S. supercomputers in development are setting out to achieve when they are completed in 2021 or later.

Overall, the latest Top500 list saw the number of supercomputers using Arm-based processors double to four this spring, three of them with Fujitsu's A64FX and the fourth with Marvell's ThunderX2.

Dominic Daninger, vice president of engineering at Nor-Tech, a Burnsville, Minn.-based HPC system builder that partners with Intel and AMD, said it could be a sign of bigger things to come for Arm-based processors, but the chip designer and its silicon partners ultimately need to ensure their processors are optimized for the right applications, supported by server vendors like Supermicro and provide an attractive total cost of ownership for greater adoption.

"A lot of things have to happen in the community, but if it's an inviting performance-per-dollar formula there, then they eventually get some attention," he said. "And if you get somebody like a Supermicro or something like that that steps in with the hardware, then it becomes pretty interesting."

New Nvidia A100 'Selene' Cluster Reaches No. 7 On The List

Fujitsu and Arm weren't the only processor vendors to hit new milestones. Nvidia debuted a new supercomputer called Selene with its new A100 GPUs that reached No. 7 on the list — a move that also brought about the first time an AMD EPYC processor landed in the top 10 systems.

Selene, which is part of Nvidia's internal research cluster, is based on Nvidia's DGX SuperPOD reference architecture and consists of 280 DGX A100 systems, 2,240 A100 GPUs, 494 Mellanox Quantum 200G InfiniBand switches, a 56 TB/s network fabric and 7 PB of high-performance all-flash storage. Each DGX A100 system is also equipped with two 64-core AMD EPYC 7742 processors.

Nvidia's GPUs were in five other top 10 supercomputers: the U.S. Department of Energy's Summit at No. 2, the DOE's Sierra at No. 3, Italian energy giant Eni's HPC5 at No. 6, Italian computing center Cineca's Marconi-100 at No. 9 and Swiss National Supercomputing Centre's Piz Daint at No. 10.

The number of supercomputers on the Top500 list with Nvidia GPUs was flat at 135, one less than the fall 2019 list. But the number of supercomputers using Nvidia's V100 GPU, the predecessor to A100, grew to 102 this year from 90 last fall. Meanwhile, Selene marked the debut of the A100 in the top 500.

Nvidia's reach in the top 500 is much larger, however, when considering its acquisition of Mellanox Technologies, which closed earlier this year. The company said between its GPUs and Mellanox interconnects, it powers two-thirds, or 333, of the world's top 500 supercomputers.

AMD EPYC Growth Continues As Intel Holds Its Ground

As a testament to Intel's continued dominance of the server market, the company's number of systems using Xeon processors was 470, the same as last fall's last, while the use of its latest Xeon generation, second-generation Xeon Scalable, more than doubled to 54.

The semiconductor giant was able to claim two of the top 10 supercomputers: Eni's HPC5 at No. 6 and the Texas Advanced Computing Center's Frontera at No. 8.

AMD, in the meantime, saw its share of top 500 systems grow to 11 from six last fall. All the growth came from the chipmaker's second-generation EPYC Rome processors, which are now represented in eight of the world's top 500 supercomputers, including Nvidia's Selene at No. 7.

AMD's Radeon data center GPUs, in the meantime, saw no growth from last fall, when there was only one supercomputer using its Radeon Instinct MI60.

Both Intel and AMD are involved in DOE supercomputer projects that are set to exceed 1 exaflop in performance. Intel's project, Aurora, is set to use the chipmaker's next-generation Xeon Sapphire Rapids CPUs and Xe Ponte Vecchio GPUs. AMD is providing next-generation EPYC processors and Radeon GPUs for supercomputer projects named Frontier and El Capitan. Frontier and Aurora are set for completion in 2021 while El Capitan has an expected completion date of 2023.

Between AMD and Arm, Intel is facing greater competition than it has for a long time, according to Daninger, the executive at Nor-Tech.

"It certainly takes a while to move in," he said. "We saw AMD four years ago when they introduced the Zen architecture, and they're gaining some foothold now."

Back to Top



    trending stories

    sponsored resources