Fujitsu to open Arm-based A64X HPC systems to public cloud

The same architecture as the top-ranked "Fugaku" supercomputer with services bundled

Fujitsu has confirmed it will be opening access to the same system architecture underpinning the global top supercomputer, "Fugaku" via an emerging cloud service.

The new Fujitsu Computing-as-a-Service (CaaS) portfolio will also provide hooks for users hoping to explore quantum simulation resources and services for AI and machine learning.

First out of the starting gate is Fujitsu Cloud Service HPC, for which the firm is taking advance orders from today, with availability expected in October.

Fujitsu said it will start serving the Japanese market first, with a global rollout to international regions including Europe, Asia Pacific, and the Americas to follow.

This will be followed by services offering access to Fujitsu's Digital Annealer technology, which uses the quantum annealing process to find the optimal solution to complex problems, as well as AI and machine learning services.

Fujitsu Cloud Service HPC includes the compute power of the Fujitsu PRIMEHPC FX1000, based on the same A64FX 64-bit Arm chip used in the Fugaku supercomputer, which still holds the crown as the world's fastest, according to the TOP500 list.

Each node in PRIMEHPC FX1000 is built around a single 48-core A64FX chip with 32GB of HBM2 memory stacked directly on top of the chip for a claimed memory bandwidth of 1,024GB/s. PRIMEHPC FX1000 supports up to 384 nodes per rack, interconnected by Fujitsu's own Tofu D high-speed fabric.

According to Fujitsu, the HPC service will include support and a suite of software and libraries to help customers deploy HPC applications. It will also offer services for performance tuning and app analysis to support customers focusing on research and analysis.

Compute node, login node, job scheduler, storage, and HPC application software will all be configured in advance, meaning that users do not have to build their own HPC environment and only need to prepare data for their analysis, Fujitsu said. Users will pay only for the resources they require.

The way subscribers will use the Fujitsu Cloud HPC service is through a monthly "HPC budget" from which compute resources will be consumed on a pay-as-you-go basis, Fujitsu said.

Fujitsu told us there will initially be three pricing plans for the HPC budget needed to use HPC nodes, depending on conditions such as storage capacity surrounding usage. These are ¥50,000 per month (approx. $400), ¥500,000 per month (approx $4,000) and ¥1 million per month (approx $8,000). These are prices for the initial service in Japan.

It appears that Fujitsu is keen to support customers without previous HPC knowledge or experience to help them benefit from its cloud service via professional operational and technical support, plus development of HPC utilization plans tailored to an individual customer's business plans.

Fujitsu also said it is aiming for a high level of compatibility between Fujitsu Cloud Service HPC and its Fugaku supercomputer to make it simple to transfer research results generated on Fugaku to CaaS services to accelerate practical applications. Conversely, Fujitsu also envisions users planning large-scale analysis and research projects to be able to migrate their applications to Fugaku if necessary.

This is a capability that scientists at Japan's RIKEN research institute are working with Fujitsu to achieve, according to the director of its Center for Computational Science, Satoshi Matsuoka.

"We are working with Fujitsu to make its CaaS highly compatible with Fugaku to support such requirements, and we expect CaaS to become an important service for quickly connecting R&D on Fugaku to industrial use and practical implementation in society," he said it a statement.

"Moving forward, we will collaborate with Fujitsu to further synergize Fugaku with this new service to provide its capabilities seamlessly in the cloud."

In this respect, Fujitsu Cloud Service HPC differs from some other HPC-as-a-service offerings such as Lenovo's TruScale and HPE's GreenLake for High Performance Computing, both of which provide physical infrastructure in the customer's own datacenter or colocation site, but as a managed service.

For enterprises, another option is to go with a cloud services provider, as discussed by our sister site The Next Platform, with AWS, Azure, and Google Cloud all offering HPC capabilities. However, these typically require the customer to put together and manage a solution themselves from individual services, while Fujitsu is offering support services as part of the deal.

Steve Conway, senior adviser for HPC Market Dynamics at Hyperion research said Fujitsu's move is not unprecedented, and there is a whole spectrum of HPC services already out there.

"Fujitsu is just basically joining the crowd," he said.

However, the bigger picture is that Fujitsu is being pushed by the Japanese government to expand availability of its supercomputer beyond the domestic market. "Japan has tried to keep up with the supercomputing leaders, and they've done a good job, but it has largely served the home market. Now word has come down from the government to open up to non-Japanese researchers," Conway said.

Fujitsu's K5 hybrid cloud service was sold and deployed as a public, private virtual or private hosted services until it waw killed off in 2018 in all parts of the world except Japan. Since then it has trained staff to be more skilled on AWS and Azure public clouds. ®