The RIKEN Institute is Japan’s largest research center, with more than 3,000 researchers working in 13 scientific centers throughout the country. Several dozen of those researchers work at the RIKEN Center for Computational Science (CCS), which currently is pursuing a three-pronged strategy across HPC, AI, and quantum computing, RIKEN CSS Director Satoshi Matsuoaka said during a webinar today.
RIKEN CCS is perhaps best known for Fugaku, the 7.6 million core supercomputer that’s currently number seven on the TOP 500 list and which held the number one spot when it debuted in the spring of 2020. RIKEN is currently in the design phase for FugakuNEXT, which it’s jointly building with Fujitsu and Nvidia and which will feature new MONAKA-X CPUs alongside Nvidia GPUs.
RIKEN CCS is designing FugakuNEXT to be a leadership-class machine that will carry the institution forward not only with classical HPC and new generative AI computing, but also with quantum computers and next generation hybrid quantum setups, Matsuoaka said in a webinar hosted today by Google Cloud Advanced Computing Community.
“We’ve been doing feasibility study of a future machine with many of our partners, not just Japan vendors but US vendors,” Matsuoaka said. “We have chosen our R&D partners and development has commenced and hopefully we have the next machine … deployed in 2029 and operational in 2030.”
The goal with FugakuNEXT is to build a Zetta-scale machine that offers an order of magnitude better performance than the original machine, within a nominal power limit of 40 megawatts, the two time ACM Gordon Bell Prize winner said.
RIKEN CCS hopes to build a new machine that can handle traditional HPC modeling and simulation workloads as easily as it can run newer AI workloads. The institute has invested substantially in initiatives to utilize AI to automate and accelerate scientific discovery, or AI for science.
At the same time, FugakuNEXT also is expected to feature prominently in RIKEN CCS’s quantum computing initiatives, particularly with hybrid quantum workloads that utilize traditional computers and quantum machines working together to tackle tough problems.
“The most complicated problems lend themselves to hybrid computing. So we [give] the exponential part to quantum, and the non-exponential speedup parts to classical,” Matsuoka said. “We’re investigating the infrastructure, how we merge these machines, how we combine them, how we schedule the resources, how do we let people program. All these issues, and what are the effective algorithms and applications.”
RIKEN CCS currently has three quantum systems as well as some quantum simulators. It has a System 2 machine with 56 superconducting qubit from IBM, a 20 qubit trapped-ion machine from Quantinuum that is in the process of being upgraded to 56 qubits, and a 64-qubit machine that RIKEN CCS built in-house along with its partner Fujitsu; the vendor could help expand this system to 100 qubits, Matsuoka said. The institute is also collaborating with other institutes that are working with other types of QC machines, including neutral atom QCs.
“So we have quite a comprehensive set of quantum machines at our disposal,” Matsuoka said. “But our real work is not to assemble these machine but to build the whole software stack and ecosystem, including software stack programing, resource scheduling, communication, etc. and also simulators.”
Getting QC to work in tandem with traditional HPC infrastructure is a big goal of RIKEN CCS. The institution has completed some early work with IBM that demonstrate the effectiveness of a quantum-HPC workflow scheduler for distributed quantum computing, or DQC. That work involved using a QC to create a sample of data for some problem, and then model it using an HPC resource, including Fugaku.
“They have been able to achieve what could be considered, not so much quantum advantage – because [one] could solve the same problem with Fugaku,” Matsuoka said. “You can use the same algorithm. But it’s much faster using quantum and hybrid.”
Riken CCS is working with industry partners like Mitsubishi Chemical, Softbank, and Toyota to investigate the potential to use hybrid quantum-HPC setups to solve their computational problems.
“In five to 10 years, when we have real machines that are DQC and have much higher number of qubits and have better scale machines on the classical side, we hope to be able to demonstrate quantum advantage over many classes of problems,” he said.
RIKEN CCS is a member of the Trillion Parameter Consortium (TPC) and is working on a host of AI for science initiatives. One of the big initiatives is to use AI to help port legacy codes for AI applications into something more modern. The institution’s investment in Nvidia GPUs and other AI accelerators with FugakuNEXT and other machines–including a 2,000 Grace Blackwell GPU cluster that will serve as a testbed for quantum-HPC work and as a development prototype for FugakuNEXT–will give it the horsepower to take on this monumental task.
“It’s one thing to do AI coding with your standard Python or Internet friendly types of codes, but we need to modernize Fortran codes and such,” Matsuoka said. “These AI coding agents… don’t work very well on arcane Fortran code. Can we use them to modernize? That will be one of the big technical challenges.”
As the design of FugakuNEXT begins, RIKEN CSS is already beginning to look to the future and a potential FugakuNEXT-NEXT machine. Matsuoka has his eyes on emerging technologies such photonic networking, advanced memory designs, and of course quantum that could help push the bar even higher.
“We’re a research institution. We want to do better than what the vendor proposed, because otherwise we wouldn’t have the rationale of being an advanced research lab,” he said. “It’s our long-term commitment to have sustainable research so that in the various cadences we have, we’ll be ready when these technologies are mature enough to be deployed en masse…The goal is to build a machine. But that’s secondary. We know the machine will be built. The question is whether it will be a really top-tier machine that people will want to replicate.”
