• SHARE

Japan’s Fugaku supercomputer — which in June ranked first in the global Top500 list of such machines, the first time for a Japanese machine in about nine years — was surprisingly not created with the aim of excelling in numerical benchmarks, unlike some of its rivals.

Instead, it was born with an “application-first philosophy,” meaning that its exclusive purpose is to dedicate its computational excellence to tackling some of the world’s biggest challenges, such as climate change, says Satoshi Matsuoka, 57, the mastermind behind the project.

“Benchmark excellence is not our priority,” he said in an interview conducted in near flawless English. Instead, he said, its success is assessed “based on how much we can accelerate the applications that are important in society.”

As the director of Riken’s Center for Computational Science, Matsuoka and his team have set out nine application areas for Fugaku to work on that are of importance to society, such as medicine, pharmacology, disaster prediction and prevention, environmental sustainability and energy.

Matsuoka began leading the team developing the next-generation supercomputer in around 2010, just before its predecessor K computer became the world’s fastest supercomputer in the Top500 benchmark by conducting more than 10 quadrillion calculations per second.

Fugaku, set to be officially launched in 2021 at Riken’s facility in Kobe, won international acclaim for becoming the world’s first supercomputer to grab the top spot in all four Top500 categories — raw computational speed, big data processing, deep learning with artificial intelligence and practical simulation calculations. It was developed jointly by the state-backed Riken research institute and Fujitsu Ltd. over a decade at a cost of ¥130 billion.

Fugaku conducted more than 442 quadrillion calculations per second in the Top500 benchmark test, which computes the machine’s raw speed. That’s nearly three times faster than the second-ranked Summit system, developed by Oak Ridge National Laboratory in the United States. Fugaku again topped the list in November.

Named after an alternative name for Mount Fuji, the supercomputer has already been used for experimental trials in various research related to the coronavirus and global weather simulations, among others, since April.

Satoshi Matsuoka, the director of Riken’s Center for Computational Science, has led the development of the world's fastest supercomputer, Fugaku, over the past decade. | COURTESY OF RIKEN
Satoshi Matsuoka, the director of Riken’s Center for Computational Science, has led the development of the world’s fastest supercomputer, Fugaku, over the past decade. | COURTESY OF RIKEN

With the K computer becoming operational in 2011, Matsuoka and a group of supercomputer researchers got together and started to consider the successor to K. But when Upper House lawmaker Renho in 2009 criticized the use of more than ¥100 billion of taxpayers’ money for K, asking what was wrong with having the world’s No. 2 supercomputer, the direction of the next supercomputer became clear: The priority was to improve the computer’s usability so that it could process various programming languages — just like any standard computer — rather than purely computational speed.

Another focus of the project was to achieve an unprecedented level of power-efficiency. Both were solved with the creation of the Fujitsu A64FX microprocessor, which runs the same programs as smartphones and PCs and at the same time is up to 20 times more powerful than its predecessor but is extremely energy efficient. Fugaku uses about 30 megawatts at full power.

About 160,000 A64FX microprocessors are used in Fugaku, and the chip is also being used by the supercomputer manufacturer Cray Inc. (a subsidiary of Hewlett Packard Enterprise Co.), marking the first time in history that a Japanese supercomputer chip has been utilized by an American supercomputer manufacturer, Matsuoka said.

“Japan re-emerging as a formidable force in the high-end semiconductor is something that we really wanted to achieve. And that’s taxpayers’ money well spent,” he said.

Here are some other excerpts from the interview with Matsuoka, who is the recipient of the 2014 Sidney Fernbach Award for his work on software systems for high-performance computing.

What kind of difficulties did you face in building Fugaku?

There were many difficulties. But one of them was that in 10 years, we had to build something that’s 100 times more power-efficient than the K computer. And that seemed like going to the moon. In the end, it turned out that we actually met our goal.

It was very hard to push the performance of the CPU (central processing unit) and push the power efficiency. One possibility had been for major CPU companies like Intel to come up with a CPU around 2020 that would satisfy the goal, but our projections quickly told us that would likely not be the case and we would fall far short of the objectives.

So just like K, it was very clear we had to build our own processor that would be much faster and much more power-efficient than a state-of-the-art CPU in 2020, and at the same time, be general purpose to be compatible with PCs and smartphones. This, of course, was an immense challenge.

What made it possible to create Fugaku?

Not just Riken and Fujitsu, but the entire Japanese HPC (high-performance computing) community came together. My center is a national center for HPC that was created along with K. That allowed us to ultimately meet our goal.

Some of the technologies in building the machine were carried over from K. That made the challenge a little easier, but the principal difficulty was building a CPU that was fast, efficient and general purpose, which K did not address. When people think about Fugaku, they think all the money was spent on hardware design and the procurement of the machine itself, but a lot of money was spent on research.

What are some of the applications Fugaku will contribute to besides the search for breakthroughs in solar and fuel cells?

There are a lot of climate models but none are really accurate, because that’s like trying to forecast the weather 50 years or 100 years ahead. And we know that the weather forecast is difficult, predicting just two weeks ahead. So we’re trying to predict the climate in many years’ time. It is very, very difficult. The demand on computing is limitless, but a machine like Fugaku will allow you to achieve much better precision with respect to what the climate may look like given the effect of the carbon dioxide emissions.

The K supercomputer became the fastest machine in the world in 2011, conducting more than 10 quadrillion calculations per second. | COURTESY OF RIKEN
The K supercomputer became the fastest machine in the world in 2011, conducting more than 10 quadrillion calculations per second. | COURTESY OF RIKEN

Are there other examples of how Fugaku can help Japan deliver on its pledge to achieve carbon neutrality by 2050?

There is work on carbon sequestration so we can trap all the carbon dioxide underground to reduce the amount of CO2 in the atmosphere. There is a lot of work in the U.S. and the work has also started in Japan, although it has not been that popular. This needs lots of simulation because we have to estimate the effect of long-term sequestration. How do we prevent that from escaping? What’s the best way of injecting CO2? Where do we inject CO2? Fugaku and other supercomputers are trying to solve these zero carbon problems in many ways.

Is your next goal creating another supercomputer that’s 100 times faster than Fugaku?

One hundred times would be extremely challenging, but we will see. Moving toward the next generation beyond Fugaku, it would be more difficult in a more fundamental way — because of the slowing down of advances in semiconductors plus many other physical limitations that we know, of course, existed. There are many challenges ahead. We have some ideas to overcome them, but we’re still early in the process. The basic research is already ongoing but we hope to be able to launch the initial project very soon, maybe next year or the year after that.

What’s the time frame for developing the next-generation supercomputer?

K was officially accepted for production in 2012. And Fugaku will be officially accepted for production in 2021. We can’t tell you when in 2021, but we’re saying as early as possible. So that’s nine years or less. We hope that we can do the next one, by the earliest, 2028 or 2029. 2028 will be a good time frame for us to at least start producing the machine, so maybe in fiscal 2029 we may see a full or partial machine. But of course the project has to be accepted first.

What are the goals you want to achieve through the use of Fugaku?

Once the advances of supercomputers stop, the advances of the rest of IT would stop as well, such that the progress of society grounds to a halt. Constant progress in information technology is extremely important to ensure the overall progress of society, and it’s not just the problem of one supercomputer. How we advance high-end computing will continuously ensure dramatic improvements in society in the years to come.

In a time of both misinformation and too much information, quality journalism is more crucial than ever.
By subscribing, you can help us get the story right.

SUBSCRIBE NOW

PHOTO GALLERY (CLICK TO ENLARGE)