Nvidia's JUPITER simulation cuts the road to useful quantum computing
NVIDIA says a 104-qubit simulation on Europe's first exascale supercomputer could help researchers design fault-tolerant quantum hardware with less overhead.
NVIDIA says it has run the largest comprehensive simulation yet of a fault-tolerant quantum computing system, using JUPITER, Europe's first exascale supercomputer, at the Jülich Supercomputing Centre in Germany. The 104-qubit simulation models a quantum processor operating with realistic hardware noise rather than in ideal conditions. 1
That distinction matters because useful quantum computers will need to keep calculating even when their fragile qubits are disturbed by errors. A clean, noiseless model can show that a technique works in principle, but a noisy model gives researchers a better indication of the resources a real machine might need.
What NVIDIA simulated
The project focused on magic state cultivation, a method used in fault-tolerant quantum computing. Magic states are specially prepared quantum resources consumed by the operations required to run useful fault-tolerant calculations. Producing them repeatedly is one of the major resource demands in a practical quantum computer.
According to NVIDIA, the approach studied in the simulation needs 90 per cent fewer physical qubits than previous methods. That does not mean a working quantum computer has suddenly been built with 90 per cent less hardware. It means the simulated technique could reduce the hardware overhead required to protect logical quantum information, if future physical implementations achieve comparable results.
The simulation covered 104 qubits and included realistic noise. NVIDIA says earlier work at Jülich modelled 50 qubits under ideal, noiseless conditions, making the new run a larger and more demanding test of how the method behaves when errors are part of the calculation.
Why JUPITER matters
JUPITER supplied the computing capacity needed to model the quantum system at this level of detail. The supercomputer uses NVIDIA GH200 Grace Hopper superchips, and the project scaled up to 8,192 GPUs after the team developed and tested its methods on NVIDIA's GB200 NVL72 system at the company's Accelerated Quantum Research Center.
The researchers used NVIDIA's cuQuantum software and CUDA-Q platform to represent and run the quantum circuits. The development work helped reduce simulations that had previously taken weeks to runs lasting days, while also allowing the team to compare results as more GPUs were added.
This is classical computing being used to study quantum computing. JUPITER did not perform the 104-qubit calculation on a functioning quantum processor, and the result is not a demonstration that a fault-tolerant quantum machine is ready for deployment. Instead, the simulation gives researchers a way to test designs and error behaviour before committing to expensive physical hardware.
NVIDIA says that kind of advance modelling could help teams assess quantum processor designs earlier, potentially reducing trial and error as they work towards applications such as drug discovery, financial modelling and materials science. Those are possible future applications, not outcomes delivered by this simulation itself.
A useful result, with a sizeable caveat
Jensen Huang, NVIDIA's founder and chief executive, said the work shows how quantum processors and GPU supercomputers can be used together to investigate problems that would otherwise be out of reach. Kristel Michielsen, director of the Jülich Supercomputing Centre, said JUPITER gives researchers a way to make better-informed decisions about designs before building them. 1
TechRadar also reports that the work used 8,192 GPUs and that the simulation incorporated hardware noise rather than relying solely on idealised conditions. 2 That context is important: the headline 90 per cent reduction describes a resource improvement in the simulated magic-state cultivation approach, not a finished quantum computer or a guaranteed reduction in the cost of a future system.
Our opinion
The strongest part of this announcement is not the promise of near-term quantum applications. It is the attempt to make the engineering problem more measurable. Quantum computing has spent years wrestling with the gap between elegant algorithms and machines that can preserve information long enough to run them. A large noisy simulation cannot close that gap on its own, but it can reveal which designs are worth building and which assumptions fail under pressure.
The next test is experimental. Researchers will need to establish how closely the simulated noise and resource requirements match the behaviour of real quantum hardware. Until then, JUPITER's achievement is best understood as a powerful design and validation tool: an important step for fault-tolerant quantum research, but not proof that useful quantum computing has arrived.