Fujitsu Receives Order From Japan’s National Institute of Advanced Industrial Science and Technology

Fujitsu Limited received an order from Japan’s National Institute of Advanced Industrial Science and Technology (AIST) for the AI Bridging Green Cloud Infrastructure (ABCI)  HPC system, which aims to accelerate AI R&D in industry, government, and academia.

The system will consist of 120 next-gen 2-way rack (4U) Fujitsu PRIMERGY GX2570 servers and a 11.2PB storage system, each accelerated by eight GPUs¹ per node and suitable for large-scale analysis using AI such as deep learning. The theoretical peak performance of half-precision floating point operations², which represents an important benchmark in the AI field, is expected to be 300 petaflops³, while that of double-precision floating-point operations⁴, which are mainly used in the conventional simulation field, is expected to be 19.3 petaflops.

The ABCI system will be added to the AI Data Center Building at the National Institute of Advanced Industrial Science and Technology Kashiwa Center in Japan with the beginning of operations scheduled for FY21.

Background
AIST launched the ABCI system on August 1, 2018, which enables advanced AI processing through coordination of algorithms, big data, and computational power, to accelerate R&D of AI technologies and their implementation in industry and other fields in society. After receiving the order to construct ABCI, Fujitsu delivered a cloud infrastructure for AI applications featuring high level of ML processing capacity, HPC capacity, and energy efficiency. Since its launch, more than 1,000 users in total, ranging from AI-related start-up companies to general electrical appliance manufacturers, have leveraged the powerful system to engage in AI R&D.

In response to rapidly increasing demand for AI R&D in industry, government, and academia, AIST will introduce a new ABCI system to bolster its existing ABCI, combining a HPC system with a large-capacity storage system while following the basic structure of the ABCI. It has awarded this order to Fujitsu in recognition of the results it delivered for the existing system, as well as the strength of its proposal, which draws on the combined technological capabilities of Fujitsu and Fujitsu Laboratories Ltd. and features a system consisting of 120 of its latest x86 servers with 8 GPUs and a 11.2PB capacity storage system.

Features of the ABCI system

• Large-Scale Cloud Platform for AI Applications with 120 Latest Servers: Fujitsu will deliver 120 of its next gen PRIMERGY GX2570 servers fueled by two 3rd Gen Intel Xeon scalable processors (codenamed Ice Lake) with Intel DL Boost per node and accelerated by eight NVIDIA A100 high-end GPUs. It anticipates that the system will provide a theoretical peak performance of half-precision floating-point operations of 850 petaflops, and that of double-precision floating-point operations is expected to be 56.7 petaflops, with the combined computing capacity of ABCI and the ABCI system. This level of performance is equivalent to second place in Japan and sixth place in the world according to the latest Top 500 HPC ranking announced in June 2020.
• High-capacity storage system for integrated operation of existing ABCI and AI Bridging Green Cloud Infrastructure system: The 11.2PB storage system built into the ABCI system can be linked with the existing ABCI storage system to enable integrated ABC and ABCI operations.

Future Plans
Fujitsu will continue to play a central role in contributing to the advancement of AI R&D in Japan through its involvement with the AIBC project with the AIST, leveraging the knowledge acquired through the construction and operation of this system for new initiatives with a variety of stakeholders in industry, government, and academia.

¹ GPUs: (Graphics Processing Units). A computing device for performing processing specialized in image processing and screen drawing.
² Half-precision floating point operations: A type of floating-point number that can be expressed on a computer in 16 bit form.
³ Petaflops: PFLOPS (Peta Floating-point Operations Per Second). Peta is 1,000 trillion (10 to the 15th power). 1,000 trillion floating point operations per second.
⁴ Double-precision floating-point arithmetic: A type of floating-point number that can be represented on a computer in 64 bit form.