Nvidia has simulated a 104-qubit fault-tolerant quantum computer on JUPITER at Germany’s Jülich Supercomputing Centre, describing it as the largest comprehensive simulation of such a system to date. Fault-tolerant machines can detect and correct errors during operation, a capability widely considered essential for practical quantum computing. Although the model represented a quantum system, the simulation itself ran entirely on conventional computing hardware.
The experiment evaluated magic state cultivation, which prepares specialized quantum states required for certain calculations. Producing these states can consume substantial hardware resources; Nvidia says the tested method requires 90% fewer qubits than earlier approaches. Crucially, the simulation incorporated realistic hardware noise, allowing researchers to assess how disturbances affect the process rather than assuming error-free operations. It builds on a November 2025 simulation at JUPITER that modeled 50 qubits under ideal, noise-free conditions.
For the latest project, researchers used Nvidia GH200 chips across 8,192 GPUs. Development began on a smaller system at Nvidia’s Accelerated Quantum Research Center, with testing expanding from four GPUs to hundreds before moving to JUPITER. According to the company, that preparation reduced simulation work from weeks to days. The team used cuQuantum software to optimize performance and CUDA-Q to model the quantum circuit. By revealing potential weaknesses before physical hardware is built, such simulations could help researchers refine designs and reduce repeated hardware testing. JUPITER, which opened in September 2025, was Europe’s first exascale supercomputer, capable of more than a billion billion calculations per second.
