JUPITER simulates a 104-qubit quantum computer, errors included

On a supercomputer in Jülich, Germany, the quantum chip does not exist. What exists is a copy of it, running on classical processors, that makes the same kinds of mistakes a real chip would make. NVIDIA reports that JUPITER, Europe's first exascale supercomputer, has completed a simulation of 104 qubits, the basic units of quantum information, with realistic hardware noise. NVIDIA describes it as the largest comprehensive simulation of a fault-tolerant quantum computing system to date.

The target of the run is a step that fault-tolerant machines cannot skip: producing magic states. These special quantum states are essential for complex calculations, but making them with low enough error rates remains a major challenge. Traditional methods need heavy hardware, and some estimates suggest that up to 95 percent of a quantum computer's physical qubits could be dedicated to producing these states, leaving relatively few for the calculations themselves.

The difference with the earlier JUPITER work is the noise. Last year the machine simulated a full 50-qubit universal quantum computer under ideal, noiseless conditions, beating the 48-qubit record set by Jülich researchers in 2019 on Japan's K computer. The new simulation models the disturbances real processors suffer during calculations. That makes it a more honest test: it shows whether a proposed method still works under practical conditions, and NVIDIA says its team used it to validate an advanced magic state cultivation approach.

The approach is different from Google's. Days before NVIDIA's announcement, Google reported that its Willow processor achieved a 40-fold reduction in errors while preparing specialized quantum states, this time on an actual quantum chip. The two efforts address different problems: simulations let researchers evaluate designs before testing them on physical hardware, while experiments like Google's measure what the hardware really does.

The project was built in stages. NVIDIA's researchers first developed their methods on a GB200 NVL72 system at the company's Accelerated Quantum Research Center, scaled from four GPUs to hundreds, then ran on 8,192 GPUs at JUPITER. The team relied on NVIDIA's cuQuantum software to optimize performance and on its CUDA-Q platform to model the fault-tolerant circuit. According to NVIDIA, this work cut simulation time from weeks to days.

For researchers, the gain is fewer blind rounds of building and testing hardware. NVIDIA says such simulations could help accelerate quantum applications in drug discovery, financial modeling and materials science, but the company gives no date for when useful quantum machines will exist, and the simulation itself does not provide one. As Jensen Huang put it, "understanding how they behave at scale is essential to building them." What is established is a more realistic test bed: a simulated machine that now carries its own errors.