Japan intends to deploy a 130-petaflops (half-precision) supercomputer by early 2018 as part of a 19.5 billion yen ($173 million) project called ABCI (for AI Bridging Cloud Infrastructure). The Ministry of Economy, Trade and Industry launched the project to boost Japan’s technical leadership on the global stage, which is seeing increased competition from China and Korea. Reuters, which broke the story, reports that Japan will outsource a portion of the digital horsepower to corporations for use in AI and big data problems.
Japan is aiming to build the system with a power envelope of less than 3 MW and a PUE of under 1.1. This would boost efficiency higher than Japan’s current flagship supercomputer, Oakforest-PACS. The Fujitsu/Intel system placed sixth on the latest TOP500 list with a LINPACK score of 13.55 petaflops (24.9 petaflops peak) and a 2.72 MW power envelope.
Japan’s National Institute of Advanced Industrial Science and Technology (AIST), the driving organization behind ABCI, plans to achieve this advanced efficiency with a combination of ambient warm liquid cooling, large Li-ion battery storage, and high-efficiency power supplies. The computer architecture will be based on commodity parts and will employ an accelerator approach.
The ABCI supercomputer will be located at the University of Tokyo’s Kashiwa Campus and will be operated as a cloud infrastructure, incorporating dynamic deployment, container-based virtualized provisioning, multi-tenant partitioning, and automatic failure recovery. It will be used for both academic and commercial R&D.
We expect Fujitsu, the vendor behind Oakforest-PACS and the coming post-K system, to bid on the ABCI tender, but it has not yet announced its intention to do so, according to Reuters. Bids will be accepted through December 8th.
While multiple nations are on track to stand up so-called “pre-exascale” machines in the next three years, the larger battle is for exascale bragging rights. The US, China, Japan and the EU have been jockeying to be the first past this golden 1,000x FLOPS post. Currently, the US is on track to stand up two “capable exascale machines” by 2023 as part of its Exascale Computing Project. China is building three exascale prototypes and intends to field an exascale machine by 2020. Japan was also working with a 2020 timeline, part of its Flagship 2020 program, but was forced to delay its ARM-based exascale supercomputing project by one-to-two years due to chip supply issues.