NVIDIA simulates 104-qubit fault-tolerant protocol on JUPITER

What NVIDIA ran, and where

NVIDIA described the run as the largest comprehensive simulation of a fault-tolerant quantum processor to date. The company said it ran on 8,192 GPUs on JUPITER at Jülich Supercomputing Centre.

NVIDIA says no earlier run modeled a quantum processor in such detail. According to a technical post from NVIDIA's CUDA-Q team, the simulation modeled coherent phase errors, which accumulate and interfere, and leakage, where a superconducting qubit leaves its two computational levels. Selected physical sites were given a third energy level, which raises the complexity of the simulation. JUPITER is Europe's first exascale supercomputer and sits at Forschungszentrum Jülich.

Why magic states decide the qubit bill

Fault-tolerant machines must keep producing magic states, a specific arrangement of qubits consumed during computation. Traditional distillation protocols spend 90% to 95% of a quantum computer's entire physical qubit budget just making the T-states that universal computation needs. Magic state cultivation replaces that overhead with a smaller, localized code-switching technique. NVIDIA says the cultivation approach it simulated requires 90% fewer physical qubits than previous methods. The company argues the gap matters because current hardware remains limited in qubit count, so any technique that lowers overhead helps.

The simulated protocol starts with a distance-3 color code, a small error-correcting code, and uses flag measurements to detect harmful faults.

It then grows the state into a distance-7 hybrid color and surface code configuration before further error-correction rounds and readout. NVIDIA's CUDA-Q technical post says the 104 qubits counted include the data and auxiliary qubits used across that sequence, rather than 104 physical qubits.

From 50 qubits to 104, and what changed

Jülich's earlier 50-qubit simulation ran under ideal, noiseless conditions and modeled a universal quantum computer. It surpassed a 48-qubit record set in 2019 on Japan's K computer by Jülich researchers and their Japanese counterparts. The 104-qubit run instead models magic state cultivation under realistic hardware noise. Simulating 50 qubits needed about 2 petabytes of memory, roughly two million gigabytes. Each extra qubit doubles the computing power and memory required, and a typical laptop can still handle around 30.

NVIDIA said the work began on an NVIDIA GB200 NVL72 system at the NVIDIA Accelerated Quantum Research Center.

There the team scaled tests from 4 GPUs to hundreds before committing workloads to JUPITER's 8,192-GPU configuration. According to NVIDIA's blog post, that staging compressed what had been weeks of simulation work into days. The company says the preparation helped the team use JUPITER's exascale capacity effectively.

What the people running it say

"Useful quantum computing will require fantastically complex, fault-tolerant systems, and understanding how they behave at scale is essential to building them," Jensen Huang, NVIDIA's founder and CEO, said in the announcement.

"Europe's investment in JUPITER gives researchers the ability to investigate questions that were previously beyond reach," Kristel Michielsen, director of the Jülich Supercomputing Centre, said in NVIDIA's announcement. She said the collaboration with NVIDIA helps scientists make better-informed decisions about the designs most likely to deliver useful applications.

What happens next

NVIDIA says the run gives researchers data to predict how future designs will perform before they are built, reducing reliance on hardware trial and error. The company points to drug discovery, financial modeling and materials science as the intended application areas. The simulation used NVIDIA's cuQuantum software and the CUDA-Q platform's fault-tolerant logical layer. No timeline has been announced for making the simulation tools publicly available to external quantum hardware partners.