Real-time decoding of surface codes on superconducting qubits below threshold with a streaming parallel blossom decoder

Measurement data from an error corrected quantum computer must be decoded at least as fast as it is generated, in order to prevent a backlog of syndrome data that grows exponentially in the T-gate depth of a computation. In this talk, we report on a real-time, streaming, correlated matching decoder integrated with distance-3 and distance-5 surface codes on superconducting qubits. Our decoder achieves throughput with a latency of 63 microseconds on distance-5 up to a million rounds, with a surface code cycle time of 1.1 microseconds. Furthermore, the logical error rate of our distance-5 surface code is suppressed by a factor of Λ=2.0 relative to distance-3 using our real-time decoder, not far from the performance of a neural network decoder that processes syndrome data offline and achieves Λ=2.18. The talk will focus on the decoding algorithm used to achieve these results: a parallelization of the sparse blossom algorithm that includes an edge re-weighting strategy to handle correlated errors.