Quantum computing is progressing towards so-called fault-tolerant systems, capable of correcting their own errors, and reducing these errors remains a central challenge. In October, Google and Nvidia each announced advances around the technique magic state cultivation, which aims to more efficiently produce the states necessary for universal quantum computations.
Qubits, the basic units of quantum computing, are very fragile: they can lose their coherence, and a system that makes too many errors is not useful. Providers therefore create logical qubits from many physical qubits, capable of detecting and correcting errors without interrupting the quantum state. However, a single logical qubit can require several thousand physical qubits, complicating scalability.
A universal quantum computer must combine Clifford gates, such as S and CNOT gates, and non-Clifford gates, such as T gates. Most error correction codes can handle the former but not the latter. Magic states, introduced in 2004 by researchers, are essential for these computations, but they are fragile, difficult to produce with high fidelity, and cannot be reused.
To compensate, magic state distillation transforms many noisy, i.e., imprecise, states into a reduced number of high-fidelity states. However, this operation can consume up to 95% of a quantum system’s capacity, creating a bottleneck. In 2024, Google researchers presented magic state cultivation: instead of producing a large number of noisy states and discarding most of them, a noisy state is placed in a small area of the chip and measurements are performed to detect errors and eliminate faulty attempts until a state with a lower error rate is obtained.
A demonstration on chip
On October 15, Google and Nvidia announced projects that, according to them, significantly advance the development of this technique. Researchers at Google Quantum AI report the first hardware implementation of the method, executed on the Willow superconducting chip. According to the company, the demonstration reduced errors by 40 times compared to uncultivated states, achieving T-states with 99.99% logical fidelity.
Less than a week later, Nvidia researchers announced they had conducted a simulation of 104 qubits on Jupiter, the first European exascale supercomputer, hosted at the Jülich Supercomputing Centre in Germany and equipped with Nvidia GH200 superchips. According to Nvidia, the method uses 90% fewer qubits than other approaches.
Sam Stanwyck, quantum product director at Nvidia, called the work of both companies a critical step toward fault-tolerant quantum computing. About Google, he said: “It is complementary to Google’s work. What Google did was a small-scale demonstration on hardware. It was a proof of concept showing that magic state cultivation actually works on real hardware.”
Google researchers, after two years of work to bring the method from theory to practice, detailed their results in an article published in the journal Science. They acknowledge the need to demonstrate the technique on a larger scale to make it realistic in a practical quantum system, while estimating that “this achievement transforms a key theoretical idea into an experimentally demonstrated reality.”
Simulating to design
According to Sam Stanwyck, simulation is essential to design the hardware and software around these approaches. “They must be simulated as deeply as possible to understand their behavior and especially how they handle errors,” he said. He also mentioned work by Amazon Web Services and Quantum Elements, with the University of Southern California and Harvard University, who also used simulation to address error correction.
Nvidia used its quantum-classical platform CUDA-Q and its software development kit cuQuantum to model hardware errors, including coherent errors and leakage. “These are very small systematic errors in quantum operations that accumulate during a computation and limit what the computer can do,” explained Sam Stanwyck. The project began on a system equipped with GB200 NVL72 at the Accelerated Quantum Research Center in Boston, before being extended to Jupiter, where tests scaled from four GPUs to several hundred, then to 8,192 GPUs.
