AI quest drives major orders for Japan’s supercomputer makers
As artificial intelligence increasingly takes center stage in computer technology, large-scale orders are heading to Japan’s supercomputer makers from research entities in need of the massive processing power those machines provide.
Fujitsu is assembling a dedicated AI-use supercomputer for the Institute of Physical and Chemical Research, better known as Riken, which plans to start using it in April.
Tokyo Institute of Technology, also known as Tokyo Tech, has ordered a large-scale system for AI-education purposes from SGI Japan, another supercomputer maker. The institute plans to introduce it in August.
And industry insiders are looking forward to an even bigger opportunity: a huge AI-use supercomputer the National Institute of Advanced Industrial Science and Technology, or AIST, wants to start using in 2018. The research body plans to invite tenders as early as July.
“There may be various large-scale supercomputer deals going forward, but everybody in the industry will focus their efforts on this one,” an executive at a manufacturer said.
The system, potentially worth over 5 billion yen ($43.3 million), is by far the big prize for makers. The institute will use a 19.5 billion yen outlay to bolster the nation’s research capacity for AI that the government set aside in its second supplementary budget for the year ending March 2017.
These orders are driven by one simple fact: The more computing power an entity has, the better placed it is in the global competition for AI supremacy. This is because deep learning, the key technology in the development of AI applications, requires massive processing power.
Deep learning consists of a computer system repeatedly processing huge amounts of data so that it can then make judgments on its own. Having extensive processing power means the computer can complete this work faster. The quicker the data is processed, the shorter the amount of time needed for deep learning to commence and run.
Japanese research organizations have recently become aggressive in seeking to raise their computing power after being alarmed by growing moves overseas to use supercomputers for AI development.
Riken last April launched a body dedicated to the research of AI, called the Center for Advanced Intelligence Project, or AIP. The new supercomputer Fujitsu is building for Riken is aimed specifically at speeding up the process to develop AI technologies using deep learning and other emerging methods, and examine their performance.
In its winning bid, Fujitsu proposed a system comprising 24 computers each installed with eight of the latest graphics processing units made by U.S. tech company Nvidia, which are designed for AI computing. Technically, the system can process data at the level of precision required for AI learning, namely, 40 quadrillion times per second on these GPUs alone.
The system will be set up at a Fujitsu data center in Yokohama, Kanagawa Prefecture, just southwest of Tokyo. AIP researchers will access it remotely.
The processing power of the system Tokyo Tech is introducing, dubbed Tsubame 3.0, will reach 47.2 quadrillion times per second. This ranks the system, which will cost Tokyo Tech about 3.7 billion yen for its construction and five-year lease, closely behind Oakforest-Pacs, Japan’s fastest supercomputer that can perform a similar task 50 quadrillion times per second.
Brought into operation last December, Oakforest-Pacs is jointly operated by the University of Tokyo and the University of Tsukuba.
Tsubame 3.0 will be made up of 540 units of a computer jointly developed by Tokyo Tech and Silicon Graphics International, the U.S. parent of SGI Japan and part of Hewlett Packard Enterprise.
It will also have 2,160 units of Nvidia GPUs, as well as 1,080 Intel central processing units.
“We’ve successfully developed a system that will have the largest number of GPUs housed in a single system in Japan, at a high degree of density,” Tokyo Tech professor Satoshi Matsuoka said. “It is also highly efficient in cooling itself.”
When completed, AIST’s AI-dedicated supercomputer will become Japan’s fastest computer, theoretically capable of performing AI-level processing at 130 quadrillion times per second.
Manufacturers hope the track record they build with research institutes will expand to businesses. For now, though, they have their sights set on participating in the AIST job, which one executive described as “one of the biggest supercomputer deals ever, and thus, not colored by a particular maker.”