NVIDIA said Oct. 19 that it and Germany's Jülich Supercomputing Centre used JUPITER, a conventional supercomputer, to model a 104-qubit magic-state preparation protocol for fault-tolerant quantum computing.
Magic states are specialized inputs for certain quantum operations, and a fault-tolerant computer must produce them repeatedly because each state is used up in an operation. The model included coherent phase errors and leakage, when a superconducting qubit slips beyond the states used to encode information. NVIDIA said studying those effects could help processor designers assess preparation methods before building hardware.
The largest configuration used 8,192 NVIDIA GH200 GPUs. NVIDIA's technical account describes an encoded state with about 35.2 trillion complex amplitudes, using an aggregate 512 tebibytes (TiB) of GPU memory, or 64 gibibytes (GiB) per accelerator. The 104-qubit count covers data and auxiliary qubits involved across the protocol. After measurements, the simulator reclaimed memory, so it did not store a full 104-qubit state at one time.
Forschungszentrum Jülich lists JUPITER's funding as 50% from the European High Performance Computing Joint Undertaking, 25% from Germany's federal government and 25% from North Rhine-Westphalia. Those shares pay for the supercomputer, not for this specific run. Jülich Supercomputing Centre director Kristel Michielsen said the investment allows researchers to investigate “questions that were previously beyond reach.”
Before the JUPITER run, the group developed and checked smaller cases on an NVIDIA GB200 NVL72 system at NVIDIA's Accelerated Quantum Research Center, using CUDA-Q and cuQuantum. Jülich's earlier 50-qubit JUPITER simulation, announced in 2025, examined a separate universal-computer task under ideal, noiseless conditions; the qubit counts do not provide a like-for-like benchmark.
Google Quantum AI separately reported on Oct. 15 that its Willow processor had been used to cultivate magic states. It said the technique still needs development at a larger scale. The Google work used a quantum processor, unlike JUPITER's conventional-computer simulation.
