Nvidia has simulated a 104-qubit quantum computer that can correct its own errors, running the job on JUPITER, the supercomputer at Germany's Jülich Supercomputing Centre. The company said it is the largest full simulation of such a machine to date.
Qubits are the basic units of quantum computers, and they are extremely sensitive to interference. Even small amounts of noise can corrupt a calculation. A fault-tolerant design is one that detects and fixes those errors as it runs, which is widely seen as the main obstacle between today's experimental machines and useful ones.
The size claim comes from Nvidia itself, and the announcement leaves out the details that would let anyone judge it. Nvidia has not disclosed which error-correcting scheme the simulated machine used, how long the run took, or what "full" means in this context.
That last point matters. In general, an exact simulation of a quantum system demands memory that doubles with every added qubit, so brute-force modeling of 104 qubits is out of reach for any computer. Simulations at this scale usually lean on shortcuts, or on circuits with structure that makes them cheaper to model. Whether this one does is the open question.
The value of such work is also general rather than specific to this announcement. Researchers use classical simulations to test error-correction designs and software before the hardware exists to run them. A larger simulated machine gives them a bigger testbed. It does not show that a physical 104-qubit machine can correct its own errors in practice, and that gap is where most of the field's difficulty still sits.