Japan plans to build a super-efficient computer that could vault it to the top of the world’s supercomputer rankings by the end of next year.
With a processing capacity of 130 petaflops, the planned computer would outperform the current world leader, China’s Sunway TaihuLight, which delivers 93 petaflops. One petaflop is one million billion floating-point operations per second.
Japan’s National Institute of Advanced Industrial Science and Technology (AIST) isn’t just aiming to build the world’s fastest supercomputers, it also wants to make one of the most efficient. It is aiming for a power consumption of under 3 megawatts — a staggering figure, given that Japan’s current highest entry in the Top500 supercomputer list, Oakforest-PACS, delivers one-tenth the performance (13.6 petaflops) for the same power.
TaihuLight consumes over 15 MW.
AIST is also aiming for a power usage effectiveness (the ratio of total power consumption, including that required for cooling, to power consumed by computing devices) of under 1.1. That’s a PUE value attained only by the world’s most efficient data centers.
The institute plans to use liquid cooling, a technique also being used by French company Atos in its supercomputer design for the French Alternative Energies and Atomic Energy Commission (CEA). Atos is aiming for a performance one exaflop (one billion billion flops) but won’t have its computer ready until 2020, while AIST expects its machine to be completed a year from now.
Where other countries have optimized their top supercomputers for calculations such as atmospheric modeling or nuclear weapon simulations, AIST is targeting machine learning and deep learning applications in the growing field of artificial intelligence (AI) with the new computer design.
The project carries the name AI Bridging Cloud Infrastructure (ABCI), and is intended for use by startups, existing industrial supercomputing users and academia, according to a document published by AIST earlier this month.