Nvidia Uses Large-Scale Simulation to Study Google’s Magic State Cultivation Method

Quantum computing research continues to advance toward practical, scalable, fault-tolerant systems, but reducing calculation errors remains a central challenge. Qubits are extremely fragile and can lose coherence easily. A quantum system that produces too many errors cannot perform useful work.

Quantum computing vendors create logical qubits from groups of physical qubits. These groups can detect, isolate, and correct errors without interrupting the underlying quantum state. Scaling remains difficult, however, because a single logical qubit may require thousands of physical qubits. Additional challenges arise in the execution of universal quantum algorithms.

Universal quantum systems must support operations already performed by classical computers as well as operations specific to quantum computing. They therefore need both Clifford gates, such as S and CNOT gates, and non-Clifford gates, such as T gates. Most error-correction codes can manage Clifford states, but non-Clifford gates remain more difficult to handle.

Researchers introduced magic gates in 2004. These gates are prepared by the computer and injected into a calculation. They are essential to universal quantum computation, but they are fragile, difficult to create with high fidelity, and cannot be reused. As a result, systems must continually produce more of them.

Magic state distillation addresses this problem by collecting many noisy or inaccurate states and converting them into fewer, high-fidelity magic states. The process is time-consuming and can consume as much as 95 percent of a quantum system’s capacity, creating a significant resource bottleneck.

In 2024, Google researchers introduced another approach called magic state cultivation. The method is intended to generate high-fidelity magic states more efficiently. Instead of producing large numbers of noisy states and discarding many of them, engineers place a noisy state in a small area of a quantum chip. They then use measurements and specialized coding techniques to identify errors and eliminate failed attempts. Over time, the process grows a lower-error magic state while reducing the hardware overhead required to run quantum algorithms.

Putting Cultivation to the Test

On October 15, Google and Nvidia announced projects focused on advancing magic state cultivation. Google Quantum AI researchers reported the first hardware implementation of the technique.

The demonstration ran on Google’s Willow superconducting quantum chip. It produced a 40-fold reduction in errors compared with uncultivated states and generated T-states with 99.99 percent logical fidelity.

Less than a week later, Nvidia researchers announced a 104-qubit simulation of magic state cultivation. The simulation ran on Jupiter, Europe’s first exascale computer, which is housed at the Jülich Supercomputing Centre in Germany and powered by Nvidia GH200 superchips. Nvidia said the approach used 90 percent fewer qubits than other methods.

Sam Stanwyck, Nvidia’s director of quantum product, described the two projects as complementary steps toward fault-tolerant quantum computing. Google’s work demonstrated that magic state cultivation can operate on real quantum hardware, while Nvidia’s simulation provides tools for studying how the method can be scaled and incorporated into future processor designs.

After two years of work to move cultivation from theory to experiment, Google scientists described their results in a paper published in Science. They noted that the technique still needs to be demonstrated at a larger scale before it becomes realistic for a practical quantum system. They also characterized the result as an experimental demonstration of a key theoretical idea.

Stanwyck said detailed simulations are needed to understand how the method behaves, particularly when exposed to errors. That information can guide the design of both hardware and software around magic state cultivation.

Nvidia is not the only organization using digital-twin simulations to study quantum error correction. Earlier in the year, scientists from Amazon Web Services and Quantum Elements, together with the University of Southern California and Harvard University, described related work in the field.

Nvidia scientists used the CUDA-Q quantum-classical platform and the cuQuantum software development kit to build detailed simulations of hardware errors, including coherent errors and leakage.

Coherent errors are small, systematic mistakes in quantum operations that accumulate during a computation and limit a system’s capabilities. Nvidia’s simulations modeled these errors and their effects on algorithms, allowing researchers to study how magic state cultivation responds. The results can help determine the requirements for quantum hardware, the construction of future systems, the code they should run, and the improvements needed in later designs.

The Nvidia team began the project on a GB200 NVL72-powered system at the company’s Accelerated Quantum Research Center in Boston. That platform was used to develop and validate the software and methods needed to run the simulations at Jupiter’s scale. The researchers expanded the tests from four GPUs to hundreds, eventually running them on 8,192 GPUs at Jupiter.

The simulations covered the path from physical behavior to application performance. By connecting errors in physical qubits with their effects on application runtime, the work allows researchers to evaluate cultivation designs, study quantum-computer architectures, and assess how errors influence useful computations.