What news from AWS re:Invent last week will have the most impact on you?
Amazon Q, an AI chatbot for explaining how AWS works.
Super-fast S3 Express storage.
New Graviton 4 processor instances.
Emily Freeman leaving AWS.
I don't use AWS, so none of this will affect me.
Cloud Services / Edge Computing / Networking

SC21: Fugaku Still Fastest Supercomputer as Exascale Looms

The latest release of the list of the fastest supercomputers in the world -- released at the Supercomputing Conference (SC 21) showed little movement for an HPC industry that is anxiously waiting for long-discussed exascale systems to come online.
Nov 17th, 2021 12:51pm by
Featued image for: SC21: Fugaku Still Fastest Supercomputer as Exascale Looms

The latest release of the list of the fastest supercomputers in the world showed little movement for an HPC industry that is anxiously waiting for long-discussed exascale systems to come online.

Japan’s massive Fugaku supercomputer retained the top spot in the biannual Top500 list of the world’s fastest systems, a position it first reached in the summer of 2020. The latest list was released this week at the start of the Supercomputing Conference (SC21). In all, the top 10 systems on the list remained fairly stagnant, although a new system – the Voyager-EUS2, built by giant cloud provider Microsoft Azure, powered by AMD’s Epyc server chips and Nvidia A100 GPUs — muscled its way into the top 10, claiming the tenth spot on the list.

The Azure system also illustrated AMD’s growing strength in high-end computing, with many of the newest systems in the top 15 being powered by the company’s silicon. The list reflects the role of AMD’s Epyc CPUs in the vendor’s return as a top-ranked chipmaker over the past four years after more than a decade of losing ground to dominant rival Intel.

It also was part of the push by the United States to place more systems on the list, eating into the lead that China had built up over the past couple of years.

Waiting for Exascale

However, in the background of all this the presence of exascale computing, a shift that will deliver systems with the compute power to run complex workloads that can’t be done — or take too long to be feasible — by current machines. Jack Dongarra — a computer science professor at the University of Tennessee, a researcher at the Oak Ridge National Laboratory and co-author of the Top500 list — reportedly said during a press conference that he had hoped to have an exascale system on this list. Unfortunately, that didn’t happen.

That said, exascale is just around the corner. The United States is expected to launch three exascale systems in the next few years. Frontier, a Hewlett Packard Enterprise system powered by AMD Epyc CPUs and Radeon Instinct MI200 GPUs, is being assembled at Oak Ridge. Aurora, another HPE system expected to come online next year at Argonne National Lab, will run on Intel Xeon chips and its upcoming Ponte Vecchio GPUs.

Intel CEO Pat Gelsinger earlier this month said that due in large part to the performance of the Ponte Vecchio GPUs, Aurora will deliver more than 2 exaflops of peak performance.

Meanwhile, El Capitan, yet another HPE system running on AMD CPUs and GPUs, is scheduled for 2023 at the Lawrence Livermore Lab.

China reportedly already has two exascale systems running, though the country didn’t submit information about either to the Top500 organization. In addition, Intel is working with chipmaker SiPearl on a joint silicon solution for the European Union’s exascale efforts.

Patrick Moorhead, principal analyst with Moor Insights and Strategy, told The New Stack that the idea that supercomputer vendors and component makers are putting a tight focus on their exascale efforts is reasonable, particularly given the heightened interest in the next level of computing. It’s also not surprising that an exascale system has yet to launch.

“Some organizations are waiting to make a huge and memorable splash” with exascale, Moorhead said. “Also, getting to exascale is hard and required more time.”

Taking a Look at the List

In the meantime, what the industry has are the systems on the Top500 list, which has Fugaku at the top with a whiff of exascale about it. The system, co-developed by Fujitsu and the Riken Center for Computational Science and powered by more than 7.6 million cores of Fujitsu’s Arm-based custom A64FX processors, had an HPL benchmark score of 442 petaflops per second, giving it more than three times the performance of the second supercomputer on the list, the IBM Power-based Summit machine.

The Top500 organization noted that in a single or further-reduced precision environment that is often used for artificial intelligence (AI) and machine learning applications, Fugaku’s peak performance was more than 1 exaflop/s, essentially in that situation making it an exascale machine.

The eight systems to follow Fugaku were in the same positions as when the list was released in June. As mentioned, the only change in the top 10 was with Azure’s Voyager-EUS2, which was spun up in the cloud provider’s east region in the United States. It is powered by 253,440 AMD Epyc chips — which come with 48 cores each and speeds of 2.4GHz — and Nvidia A100 GPUs with 80GB of memory. It has a performance of 30.05 plops/s.

The relatively unchanging list comes as “funding across organizations on a percentage basis has stayed relatively constant year to year,” Moorhead said. “There’s only a breakthrough when a different approach is taken or there are architectural changes.”

Most Interconnects are Ethernet or InfiniBand

That includes the interconnects used in the system as well. The Top500 organization noted that Ethernet was still the dominant interconnect, being used in 240 machines. InfiniBand accounted for 180 supercomputers while Omni-Path, a technology developed by Intel and spun out by the chipmaker last year (it’s now owned by Cornelis Networks) was in 40 systems. Fugaku uses the Fujitsu-developed Tofu D interconnect. Until the Intel-developed CXL technology gains traction, Moorhead said he doesn’t expect to see much movement beyond Ethernet and InfiniBand in the interconnect space.

In the supercomputing competition between nations, the United States gained some ground on China, the two countries that dominate the Top500 list. China saw its total drop from 186 systems to 173, while the United States jumped from 123 to 150. Combined, the two countries account for almost two-thirds of the systems on the list.

Group Created with Sketch.
TNS owner Insight Partners is an investor in: The New Stack.
THE NEW STACK UPDATE A newsletter digest of the week’s most important stories & analyses.