Fielding a leadership-class supercomputer doesn’t happen overnight or even in the span of a year; rather it is a multi-year undertaking. Development of the record-setting, Arm-based Riken-Fujitsu “Fugaku” system started in 2014 under Japan’s “Flagship 2020 project,” managed by Riken. As director of the Flagship 2020 project, Dr. Yutaka Ishikawa pioneered development of Fugaku (previously known as “post-K computer”).

In June 2020, Fugaku became the world’s number-one supercomputer, setting a new Top500 record with 415.5 Linpack petaflops of performance. Fugaku swept a number of other HPC benchmarks as well, and is one of the world’s most energy-efficient supercomputers (placing within the top ten of the Green500).

In this interview with HPCwire’s Tiffany Trader, Dr. Ishikawa shares Fugaku’s origin story, describing the project’s goals and how they were achieved, leading to a balanced supercomputing system that provides up to 100x speedup over the K-computer for target applications. Fugaku was put into service in June 2020, a year early to help with the fight against COVID-19.

“Our goal was not to get the number one rankings on those benchmarks. Our goal is to provide the world’s highest-level supercomputing environment to a wide range of application fields. To do that we have taken a a co-design approach together with application developers,” said Dr. Ishikawa.

The work is guided by the Japanese proverb: onaji kama no meshi wo kuu. The literal translation is “to eat rice out of a shared pot,” roughly meaning to live under one roof.  To Dr. Ishikawa, the phrase signifies “working together, sharing the experiences, solving issues together and sharing in the successes.” This is how Dr. Ishikawa views the relationship between Riken and Fujitsu. “[Although] we are client and contractor, we run the project together, sharing the experience and happiness,” he said.

“The biggest challenge of the project,” said Dr. Ishikawa, “was the prediction of long-range hardware technologies.” The budget and the performance goal was determined in 2013, but availability of the expected technology, notably 10nm server silicon, was delayed, he said. Riken had to extend the deadline of the delivery, develop and acquire next-generation silicon technology with Fujitsu, and then work on shortening the delay as much as possible. The other challenge was supporting half-precision floating point operations, used for AI workloads, added Dr. Ishikawa.

Watch the full interview below: