AsianScientist (Aug. 15, 2018) – If you’re a small tech company running some sort of deep learning application—whether it’s medical diagnostics, price forecasting or image or voice recognition—chances are you’re also looking to scale that application up. To do so, you’ll need to beef up your company’s computational muscle.
But as things stand now, your choices are limited. You could build your own cluster, but that would cost an arm and a leg to procure and maintain. Or you could tap into computing services offered by the likes of Amazon and Google—but that would mean ceding some measure of control over your data and computational resources to the whims of big tech.
Soon, however, you might be able to avail yourself of a third option: a powerful new artificial intelligence (AI) supercomputer known as the Artificial Intelligence Bridging Cloud Infrastructure (ABCI), built by Fujitsu for Japan’s National Institute of Advanced Industrial Science and Technology (AIST).
Scheduled to launch by end 2018, the ABCI will offer industry and academic researchers cloud access to compute and storage capacity that is particularly suited for AI applications, said AIST vice president and ABCI project lead Dr. Satoshi Sekiguchi.
“Big data is getting bigger. Even if you have your own small cluster, it takes a month just to complete one round of training a network. That is not realistic; for research to be competitive, we need infrastructure that reduces the turnaround time of each job. We need a platform where advanced researchers can come together,” Sekiguchi told Supercomputing Asia.
Breaking rank
When news of the ABCI broke in 2016, media speculation was rife that the Japanese were building a supercomputer that would topple China’s Sunway TaihuLight from its perch atop the TOP500 ranking.
But this was not the intent, said Sekiguchi, adding that AI supercomputers like the ABCI play in a different arena. The TOP500 ranking is largely determined by how fast supercomputers can crunch double-precision (64-bit) floating point operations. But unlike traditional high performance computing, deep learning applications don’t require such high precision—for training a network, half precision (16-bit) will suit you just fine.
Designed to maximize these less precise operations, the ABCI now clocks in at 550 half-precision petaFLOPS or ‘AI-FLOPS’—an order of magnitude faster than the 47-peta-AI-FLOPS TSUBAME 3.0, its AI supercomputer brethren at the Tokyo Institute of Technology (TiTech).
“We haven’t decided if we will submit our numbers to the TOP500, because our primary focus is on half precision rather than double precision at this moment,” said Sekiguchi.
That said, the ABCI is not too shabby even by TOP500 standards—its 37 petaFLOPS of double-precision speed would put it in the top spot in Japan and third globally, assuming the official LINPACK benchmark is run on it in the future.
AI for all
A widely respected high-performance computing pioneer with a background in parallel computing, Sekiguchi has spent his career building grid- and cluster-computing infrastructure for Japan’s research community, as well as developing advanced highspeed networking that has connected Japanese researchers with their international counterparts.
In 1997, for example, Sekiguchi was involved in setting up the Ninf network infrastructure, which allowed researchers to access data, hardware and software across a distributed computing environment—a project he remembers as particularly challenging as the team had to build everything, including security management and application programming interfaces (APIs), from scratch.
“But this kind of distributed software development was very exciting for us. We found good partners for taking care of the bottom of the stack, such as security and discovery. By working very closely with these other teams, we were able to take off the lower layer of our system,” he recalled.
In many ways, the ABCI is a natural extension of Sekiguchi’s careerlong efforts at making connections and fostering collaborations. By providing access to AI computing on the cloud, the machine fills the gaping hole in computational resources that currently hampers closer collaboration between academics, companies and vendors of IT solutions, as well as narrows the widening chasm between small players and large companies with deep pockets, said Sekiguchi.
“Some companies have their own data, but not the skill [to analyze it] and understand what is happening. They might want to work with IT vendors, but IT vendors may not have the resources to support this kind of activity,” he explained. “This is an opportunity for us to accelerate open innovation.” While not yet finalized, the ABCI’s prices will be cheaper than those offered by commercial cloud service providers, as the project does not intend to turn a profit, added Sekiguchi.
Companies from the manufacturing, autonomous vehicle and heavy equipment industries have already registered interest, as have academic researchers in areas as diverse as genomics, speech recognition and particle physics, said Sekiguchi. All users, both within the country and internationally, will be able to tap into the ABCI via highspeed networks such as Japan’s 100-gigabit SINET5, which has links to Asia, Europe and the US.
The greening of AI
Aside from the promise of delivering AI capabilities over the cloud, another unique aspect of the ABCI is its high energy efficiency, said Sekiguchi. The data center, located on the Kashiwa campus of the University of Tokyo, will be chilled by a hybrid water-air cooling mechanism, designed with the help of green supercomputing pioneer Professor Satoshi Matsuoka of TiTech.
Since liquid conducts heat away more efficiently than air, this combination uses much less power compared to a fully air-cooled system. Further, the starting temperature of the water and air coolants can be as high as 32 and 35 degrees Celsius respectively—higher than the temperature of the ambient air except in the hottest months of the year—thus eliminating the need to use extra energy to cool them, said Sekiguchi.
Like most buildings in Japan, the ABCI data center was designed with earthquakes in mind. But rather than shelling out a fortune to install antishaking dampeners on a multi-storey building, AIST decided to organize the data center across a single storey, atop a slab of hard concrete. This simple design took a mere month to build, said Sekiguchi.
Cutting the fab
In addition to acting as a community resource, the ABCI also serves as a test of the concept of ‘AI fabrication,’ or ‘AI fab’ for short, said Sekiguchi. Just as chipmakers like Qualcomm farm out the expensive and finicky process of semiconductor fabrication to specialized facilities, researchers can now leave the AI fab (the building and maintenance of AI computational resources) to the ABCI—that is, they can go ‘fab-less,’ said Sekiguchi.
“At the end, the user receives the trained network, so they don’t need to go through the whole process of building it and maintaining the computational resources needed to do so,” said Sekiguchi. “This is deep learning for the future.”
The idea of an AI-dedicated computational platform has caught the attention of other countries in the region, including South Korea, Singapore, Taiwan and Thailand, which are now starting to explore similar infrastructure, said Sekiguchi. He thinks that by making AI computing more affordable and accessible to smaller players, business models like the ABCI’s can offer Asia a competitive edge over US tech giants, which now mostly cater to consumers or to other large businesses.
“[Small businesses] are kind of a niche right now, but the market is growing because small and medium enterprises really need to deploy AI to improve their business; yet they don’t want to give their data to Amazon or Google. But if they have access to computing resources that they have control over, then there is a good opportunity for them to make things happen.”
Sekiguchi’s top priority over the next two years, he said, is to organize an ecosystem of industry and academic users around the ABCI, a task he views as more important than the addition of more computational resources.
“The AI business we’re focusing on isn’t a single-capability business; it’s more like a capacity offering, a system that people can use.”
This article was first published in the print version of Supercomputing Asia, July 2018.
Click here to subscribe to Asian Scientist Magazine in print.
———
Copyright: Asian Scientist Magazine; Photo: Shutterstock.
Disclaimer: This article does not necessarily reflect the views of AsianScientist or its staff.
