Alice & Bob decodes LDPC-cat error correction on GPUs through NVIDIA NVQLink
Short answer: French cat-qubit company Alice & Bob has decoded its LDPC-cat error-correcting codes on GPUs, using NVIDIA's NVQLink interconnect and Quantum Machines' OPX1000 controller, according to a Quantum Machines blog post this month. The test used synthetic data, not a live quantum processor.
In the Quantum Machines post, the controller maker explains why decoding matters: error correction repeatedly measures syndromes, and a classical decoder must turn them into corrections fast enough to avoid a backlog. For realistic superconducting designs, it says controller-decoder loops need to stay in the tens of microseconds. Its Open Acceleration Stack links the OPX1000 to GPUs, CPUs and FPGAs with a 2 to 4 microsecond round trip, is NVQLink-compliant and supports CUDA-Q, CUDA, C++ and Python.
LDPC-cat codes pair cat qubits, which suppress one type of error in hardware, with two-dimensional low-density parity-check codes, with the aim of cutting the overhead of fault tolerance. Quantum Machines says Alice & Bob reports this as the first time such codes have been decoded on GPUs through NVQLink.
The team also found that batching helps. Quantum Zeitgeist reported on 11 November that decoding many syndromes together made latency grow sublinearly, so the time per syndrome falls as the batch grows, which suggests GPU efficiency improves as machines produce more syndrome data.
Both accounts say the next step is real-time decoding with live data, which will require the controller-to-GPU path to keep pace with the machine's cycle time.
