Nvidia and researchers at Germany’s Jülich Supercomputing Centre have used 8,192 GPUs to simulate a 104-qubit process for fault-tolerant quantum computing on JUPITER, Europe’s first exascale supercomputer.

Nvidia described the experiment as the largest comprehensive simulation of a fault-tolerant quantum processor to date. The simulation focused on magic-state cultivation, a technique for producing the quantum states needed for certain complex quantum calculations. The process starts with an error-prone state, checks it for faults and builds additional protection against errors. Reliably producing these states can require considerable hardware, making this a major challenge for fault-tolerant quantum computing. Nvidia claims this approach could cut the number of physical qubits needed to produce magic states by up to 90% compared with earlier methods.

Reducing the number of qubits, however, is only one part of making the process practical. At a press briefing, Sam Stanwyck, Nvidia’s director of quantum product, explained that researchers must also determine how quickly magic states can be produced and whether that rate can support the calculations a quantum computer needs to perform. “You may find a way to use fewer qubits, but it takes longer to produce each state,” he said.

Understanding those trade-offs also requires accounting for errors in the underlying quantum hardware. That was a key advance in the JUPITER simulation, which modeled coherent phase errors that can accumulate and interfere with quantum operations, as well as leakage, which occurs when a qubit moves outside its intended computational states. An earlier Jülich simulation modeled 50 qubits under ideal, noise-free conditions. By incorporating realistic error sources, the new simulation allows researchers to evaluate how magic-state cultivation might perform under conditions closer to those of physical quantum hardware.

Running such a detailed simulation required substantial computing resources. The simulation used 512 tebibytes of GPU memory, or more than half a petabyte, to store the quantum-state data. According to Nvidia’s technical account, researchers used several techniques to reduce memory requirements, including reusing storage as qubits were measured and released.

In addition to GPU capacity, the record also required specialized software, Stanwyck explained. Researchers used CUDA-Q Logical to describe the steps involved in magic-state cultivation and experiment with the circuit’s design, while cuQuantum provided the computing tools needed to execute the simulation across the thousands of GPUs.

The team also developed and tested its methods on a GB200 NVL72 system at the NVIDIA Accelerated Quantum Research Center, working with configurations ranging from a handful of GPUs to hundreds while larger simulations ran on JUPITER.

The JUPITER experiment also highlights the computing resources available at Jülich and Europe’s investment in quantum computing infrastructure. The supercomputer was procured through the EuroHPC Joint Undertaking, which is also funding other efforts to integrate quantum processors with conventional supercomputers. Jülich operates its own quantum computing facility, JUNIQ, which provides researchers access to quantum hardware integrated with its HPC infrastructure.

Stanwyck expects the need for classical simulation to grow as quantum hardware advances. Researchers will require “more and more simulation and more detailed simulation, not less,” he said.