Summary

• NVIDIA and Jülich simulated a magic-state preparation protocol involving 104 physical qubits using 8,192 GH200 GPUs on the JUPITER supercomputer.
• The simulation incorporated coherent phase errors and leakage into additional superconducting qubit energy levels, rather than relying solely on simplified random error models.
• NVIDIA says magic-state cultivation can reduce physical qubit overhead by around 90% compared with conventional approaches, but the result is a classical simulation, not an operating 104-qubit quantum computer.

NVIDIA and researchers at Germany’s Jülich Supercomputing Centre have used 8,192 graphics processors to simulate a quantum error correction protocol involving 104 physical qubits, examining how hardware imperfections could affect the design of future fault-tolerant quantum computers.

The research, announced on 19 October, was conducted on JUPITER, Europe’s first exascale supercomputer, using NVIDIA’s CUDA-Q and cuQuantum software. The simulation examined magic-state cultivation, a technique intended to reduce the resources required to perform certain operations reliably on an error-corrected quantum processor.

The largest calculation involved 8,192 NVIDIA GH200 GPUs and incorporated error mechanisms associated with superconducting quantum hardware, including coherent phase errors and leakage beyond the two energy levels normally used to represent a qubit.

NVIDIA describes the work as the largest comprehensive simulation of its kind and says magic-state cultivation could reduce physical qubit requirements by approximately 90% compared with conventional methods. The comparison concerns resources required for particular fault-tolerant operations, rather than the complete hardware requirements of a finished quantum computer.

Reliable operations require more than protected memory

Quantum processors use qubits to store and manipulate information, but the physical systems representing that information are sensitive to disturbances. Imperfect control signals, unwanted interactions and environmental noise can change quantum states and introduce errors during calculations.

Error correction addresses this problem by distributing information across several physical qubits and repeatedly checking for signs of faults. Under suitable conditions, the resulting logical qubit can preserve information more reliably than its individual physical components.

Protecting stored information is only part of the requirement. A useful fault-tolerant processor must also perform operations that allow sufficiently varied quantum algorithms to run, including operations that are difficult to implement using the basic techniques employed in many error-correcting codes.

Magic states provide a resource for carrying out those additional operations. Their preparation requires a process capable of detecting and suppressing errors while producing states of sufficient quality for subsequent computation.

Conventional magic-state distillation can consume substantial computing resources because it uses multiple imperfect states to produce fewer, more reliable ones. The physical qubits allocated to this work add to the hardware required for the main calculation.

Magic-state cultivation offers a different approach. It begins with a state prepared in a relatively small error-correcting code, checks for faults and progressively increases the code’s protection. The aim is to obtain a sufficiently reliable state with less supporting hardware.

The potential saving is relevant to processor design because every physical qubit requires control, measurement and supporting infrastructure. Reducing the number needed for particular operations could simplify future architectures, provided the method performs as predicted on actual devices.

JUPITER examines errors that simplified models miss

The simulated protocol begins with a distance-three colour code and grows into a distance-seven hybrid colour and surface-code configuration. Additional measurements are used to identify potentially harmful errors as the procedure develops.

Its 104-qubit count includes the physical data qubits and auxiliary qubits involved in the simulated sequence. It should therefore not be interpreted as 104 independently usable logical qubits.

Researchers included coherent phase errors, which can accumulate and interfere through successive operations. They also modelled leakage, which occurs when a superconducting device moves outside the two states selected for computation.

Leakage requires particular attention because superconducting circuits have physical energy levels beyond the computational pair. A device entering one of those levels can behave differently from an ordinary bit-flip error, affecting later operations and measurements.

Selected sites in the simulation were consequently represented with a third energy level. That increased the computational complexity while allowing researchers to examine error behaviour that simpler two-level models would not reproduce.

The simulation retained the non-Clifford operations involved in preparing the magic state. Those operations are important because some restricted classes of quantum circuits can be simulated relatively efficiently, whereas more general calculations are substantially harder to model.

Development and validation work also took place at the NVIDIA Accelerated Quantum Research Center before and alongside execution on JUPITER. This arrangement allowed researchers to refine their simulation methods before running the largest calculations on the supercomputer.

The results provide evidence about the behaviour of the selected protocol under the specified noise conditions. Their predictive value depends on how closely those assumptions represent the physical processors eventually built.

Classical computing helps select quantum designs

Quantum hardware development involves decisions about qubit layout, control operations, measurement sequences and error correction architecture. Changing those arrangements after manufacturing can be expensive, particularly where the processor is integrated with specialised cryogenic and electronic systems.

Detailed simulation allows engineers to compare proposed designs and investigate failure mechanisms before committing to physical implementations. It can also identify which errors have the greatest influence on a protocol’s performance.

However, modelling realistic quantum systems becomes computationally demanding as the number of interacting qubits increases. The representation of quantum states can grow extremely quickly, while particular algorithms and mathematical structures determine which problems remain tractable on classical computers.

The use of thousands of GPUs for the Jülich experiment illustrates that resource requirement. It also demonstrates a commercial role for conventional accelerated computing within quantum research, alongside the development of actual quantum processing hardware.

NVIDIA supplies technology used in both parts of that process, through conventional computing systems and software designed to support quantum simulation and hybrid applications. The research could inform future processor design, although its commercial significance will depend on whether the methods lead to measurable improvements in functioning hardware.

Techopia has separately covered Quantinuum’s experiments reducing logical quantum errors. Those results concern measurements from a physical quantum processor, whereas the Jülich work investigates a proposed fault-tolerant procedure using classical simulation.

The distinction also applies to NVIDIA’s estimate of a 90% reduction in physical qubit overhead. It describes an advantage associated with the investigated cultivation approach and its comparison with other preparation methods, not a verified reduction in the size or cost of a complete commercial machine.

Hardware experiments will test the predictions

To establish whether the simulated advantages can be reproduced, researchers will need to implement the relevant operations on physical quantum hardware and measure their performance under real operating conditions.

Those experiments must account for variations in qubit quality, calibration, control electronics and interactions that may be difficult to capture completely in a numerical model.

The resulting evidence could influence decisions about which error correction architectures deserve further investment and how much supporting equipment future processors require.

NVIDIA and Jülich have demonstrated a way to examine one demanding fault-tolerant operation at substantial computational scale. The next technical milestone is agreement between those simulated predictions and experiments on quantum processors capable of running comparable procedures.