RIKEN turns on ROQUO, a hybrid quantum-classical supercomputer in Kobe

RIKEN's new ROQUO supercomputer in Kobe began operations on June 19 with a job it was built to do: physically link a Quantinuum Reimei trapped-ion system, an IBM Quantum System Two, 540 NVIDIA Blackwell GPUs across 135 GB200 NVL4 nodes, and Japan's Fugaku flagship into a single control plane. With that wiring, a partnership the U.S. and Japan announced on paper now has an exercisable testbed for the unglamorous work of circuit synthesis, low-latency classical-quantum sync, and real-time error decoding, the kind of plumbing that hybrid quantum-HPC integration will live or die on.

ROQUO was activated at the RIKEN Center for Computational Science under Japan's JHPC-quantum project. On the classical side, 540 Blackwell GPUs sit across 135 GB200 NVL4 compute nodes, where a GB200 NVL4 pairs two Blackwell GPUs with one Grace CPU in a rack-scale package built for tightly coupled training and HPC work. The nodes tie together with NVIDIA's Quantum-X800 InfiniBand and measured 19.80 PFLOPS FP64 on the HPL LINPACK benchmark, the standard HPC throughput test. The Quantum-X800 plus RIKEN's software-based SQC Interface running on NVIDIA's CUDA-Q platform is what the lab calls the unified control plane, a way to keep classical GPU arrays, Fugaku, and the attached quantum processing units (QPUs) on the same clock for jobs that have to ping-pong between them in microseconds.

What gets attached matters as much as the GPU count. The quantum backends are Quantinuum's Reimei trapped-ion system, installed on the RIKEN campus, and an IBM Quantum System Two, branded ibm_kobe. Trapped-ion QPUs store information in the internal states of charged atoms held in electromagnetic traps; IBM's System Two is the company's modular superconducting platform. Putting both in the same roofline as a Blackwell cluster, on a dedicated interconnect with a sync layer, lets the lab ask whether some research workloads benefit from quantum coprocessing instead of running on classical HPC alone.

The published workloads are still research-stage. RIKEN, Quantinuum, and NVIDIA are running evolutionary AI circuit synthesis on CUDA-Q, where the lab auto-generates and optimizes native quantum circuits for the Reimei QPU rather than hand-designing them. RIKEN's SQC Interface handles the low-latency handshakes between GPU arrays, Fugaku, and the QPUs. NVIDIA's open Ising AI models, where Ising is a 20th-century statistical-mechanics model of interacting spins that has been repurposed to map onto QPU calibration and error-decoding problems, are being used for automated calibration, state preparation, and real-time quantum error correction (QEC), the housekeeping that decides whether a quantum result is trustworthy. The first application area the lab highlights is quantum chemistry, where the hybrid stack would accelerate molecular-simulation workloads that classical HPC alone struggles with.

The framing is "marketed as" among the world's first full-scale operational platforms deploying the NVIDIA GB200 NVL4 architecture, per Quantum Computing Report's coverage, and the lab is using the deployment as the testbed for a U.S.-Japan Genesis Mission pipeline. The source basis describes a hybrid testbed that has just turned on, with three research-stage plumbing workloads, plus a chemistry demo track, now exercisable, and no published result that classical HPC alone could not match. The stake is that the U.S. and Japan now share a working classical-quantum loop they can iterate on, and that the plumbing choices made here, CUDA-Q, the SQC Interface, Quantum-X800, will shape what hybrid quantum-HPC integration looks like at the next lab that tries it.

RIKEN R-CCS is publishing operational updates here. The Quantum Insider's secondary write-up lays out the same components in more detail. Quantum Computing Report's follow-up piece on the integration work tracks the SQC Interface roadmap.