Back in the day, supercomputers were built for a specific purpose, usually deep research and modeling, and run for 4-5 years before being replaced by a new design but for another specific purpose.
But thanks to rapid advances in microprocessor architecture, cost and energy pressures and a demand for more flexible compute in high-performance compute (HPC), supercomputing is entering a new era. One of the biggest and most stunning milestones in that shift emerged in mid-June when the Top 500 supercomputer rankings were released, and No. 1 in the world was Japan’s Fugaku Arm-based supercomputer. The top ranking and the energy-efficient engineering achievement signal that an era of homogeneous design is fast coming to a close.
Fugaku (an alternative name for the Japan’s Mt. Fuji) was installed in the spring of this year at the RIKEN Center for Computational Science in Kobe. It achieves 416 Petaflop/s on the HPL benchmark easily exceeding the old #1 machine by 2.8x. In single or half precision (16-bit) tasks, which are often used in machine learning and AI applications, its peak performance is actually above 1,000 PFlop/s (= 1 Exaflop/s) and because of this, it is often introduced as the first ‘exascale’ supercomputer, according to the Top 500 organization.
Its high-bandwidth memory (HBM) is 32 GB per node at approximately 1 TB/sec of bandwidth per socket, for a total of 4.85 PB of memory capacity and 163 PB/second aggregate memory bandwidth.
The machine, which runs Red Hat Enterprise Linux unmodified, contains 158,976 Fujitsu A64FX processors (each containing 48 Arm cores for a total of more than 7.5 million Arm cores) running at 2.2GHz spread over 432 racks.
Just as impressive is Fugaku’s power efficiency. The machine ranked ninth in the world in June in the Green 500 supercomputing rankings, coming in at 14.665 GFlops/watt. This is notable since compute efficiency is fast becoming an important metric of success as silicon-level trends like Dennard scaling slow markedly.
Professor Satoshi Matsuoka, who helped lead development of Fugaku and has been deeply involved in energy efficient computing for more than a decade, said, “We hope that the chip we developed for this machine or what we design for future generations of machines makes it into mainstream market and sells into the millions.”
The foundation for that chip is the Arm Neoverse platform, a technology introduced in 2018 to meet the unique requirements of HPC and data center compute architects. It was also an acknowledgement that the days of homogeneous one-size-fits-all, server farms powered by a single, general-purpose compute architecture are being displaced by solutions that allow greater vendor choice and flexibility in how to distribute processing at optimal points along the compute spectrum from cloud to edge.
Since then, the Arm ecosystem has gained considerable traction in the HPC and hyperscale segments of computing. Consider just a few of many milestones:
- AWS Graviton2 GA – Amazon EC2 M6g instances were the first-to-market using the Arm Neoverse N1 platform and deliver up to 40% better price performance over the previous x86-based gen instances at a 20% lower price. C6g and R6g recently became GA, offering compute- or memory-optimized tuning options.
- Sandia National Laboratories began installing the first Fujitsu PRIMEHPC FX700 with Arm-powered Fujitsu A64FX processors. This system is the first that closely couples efficient and powerful Arm processors to really fast memory to help break down the memory-speed bottleneck.
- SiPearl signs with Arm – In April, SiPearl announced it signed a major licensing agreement with Arm for the development of its first-generation microprocessors based on ‘Zeus.’ Backed by the European Processor Initiative, SiPearl intends to develop Europe’s first exascale-class supercomputer.
- Ampere Altra – Ampere announced Ampere Altra, the industry’s first 80-core server processor which is based on the Arm Neoverse N1 platform. Ampere Altra will deliver the performance-per-watt, flexibility, and scalability needed to address a diverse suite of compute-intensive cloud applications to edge analytics.
- Marvell ThunderX3 – Marvell announced the ThunderX3, featuring up to 96 custom Arm v8.3 cores with SMT4 capability. Marvell ThunderX3 is designed to deliver high throughput and performance leadership on cloud workloads while reducing TCO through power efficiency.
- NVIDIA has announced full CUDA 11 support for Arm, which enables a new path to designing and deploying highly energy-efficient, AI-enabled exascale compute.
This momentum is supported by a widening software ecosystem as well as continual updates to the Arm developer program that make it easier to build applications on Arm. Arm is driving the development and deployment of cloud-native applications, focusing heavily on open source and standards for compatibility, ease of adoption, and flexibility in vendor choice. We offer the software, tools, technical documentation, and support required to design and build with Arm-based solutions and get to market quickly.
For Matsuoka, the top supercomputer ranking is important recognition for a decade of hard work to design highly efficient, cost-effective machines. But it’s also crucial that the technology has a knock-on effect that benefits other key areas of computing.
“We helped advance the state of the art in IT,” he said. “That’s why Neoverse is important. You need powerful, efficient machines in the data center.”
Highly efficient supercomputing resources are right now being leveraged by scientists keen to model drug interactions in the fight to find a cure for COVID-19. They’re also being used to build digital models of the physical world – digital twins, as Matsuoka puts it – that will enable scientists to better solve problems that face us today and avoid those we anticipate tomorrow.
“To make the cyber physical concept work, you need powerful capabilities to recreate the universe in cyberspace. You need big machines to do that,” he said. “All the supercomputing technologies are good, but we have to demonstrate our value in terms of demonstrating compute efficiency.”
For more information, please visit Arm’s infrastructure solutions page.