IBM patent filing describes sharing quantum processors by reusing noise models
IBM is seeking patent protection for a method that would let several users share one quantum processor while reducing the work needed for error mitigation, The Quantum Insider reported on November 9. The application is titled "Quantum Queue Optimization for Error Mitigation." The Quantum Insider says it was filed on May 5, 2025.
The filing describes a proposed system. It does not report measured gains in throughput, waiting times or cost.
The problem it targets
According to the report, the application addresses two sources of waste. First, a program may need only part of a processor's qubits, but exclusive access leaves the other qubits idle. Second, some error-mitigation methods need extra experiments to learn how noise affects particular groups of gates, and that work gets repeated.
Running jobs side by side can cause crosstalk, which changes the errors that have to be accounted for. IBM's approach builds those effects into the decision about which programs to run together.
How the scheduler would work
- Reuse noise models: the system would store noise models, which describe the errors tied to specific operations, and look for queued programs whose combined operations match a model it already has. - Match connectivity: a model applies only to the same set of active qubit connections. A model measured for one pattern does not automatically cover another, even a subset. - Keep models fresh: because noise drifts, the system would update frequently used models and drop stale ones. It could move a queued job forward so it runs while a matching model is still valid. - Fill gaps: if one job finishes on its qubits while another is still running, more work could be fitted into the opening without lengthening the run.
The filing gives an example on a 127-qubit processor. One program uses 52 qubits and another uses 30, with four unused qubits between them to limit crosstalk.
Policies, users and billing
The application describes three sharing policies. The strictest combines jobs only when the needed noise models already exist. A second allows new measurements and saves the resulting models. A third derives the needed noise information from existing models. The choice could depend on queue length, available resources, interference and job priority. Machine learning could help decide where programs run.
Users would control whether and how their jobs are shared, and the billing system could offer rebates to encourage participation. The filing lists interference, possible information leakage between users and coordination difficulty as challenges.
The inventors listed, according to Patentlyze as cited by The Quantum Insider, are Ana Paula Appel, Daniel Josef Egger, Kaelyn Jae Ferris, Sanjay Kumar Lalta Prasad Vishwakarma and Razvan Paul Sebastian Brandhofer. The Quantum Insider has previously covered Columbia Engineering's HyperQ, a quantum virtual machine approach to sharing processors that was tested on IBM hardware.
