Parallelized Stabilizer Simulation for QEC With Efficient non-Clifford Noise Modeling

We present a GPU-parallelized stabilizer simulator to improve scaling for QEC workloads, exceeding the state of the art per-gate simulation time within a few hundred qubits. Our simulator efficiently supports general Clifford error channels beyond the standard Pauli errors, and general non-Clifford errors via stabilizer channel decomposition. Using a quasi-probabilistic mixture of Clifford and Reset operations, we enable accurate and fast simulation of non-Clifford channels by taking advantage of the massive parallelism available on graphics processors. Of particular interest, we notice that compositions of Clifford and non-Clifford channels can reduce the negativity in quasi-probabistic decompositions, which can entirely eliminate exponential time overhead in certain regimes. We specifically show that qubit T1 relaxation with T2 decoherence can become a completely positive compound probability distribution for many types of qubits. We model previously impractical surface code and qLDPC code performance with exact T1/T2 channels, and show that logical error rates and code thresholds become significantly more optimistic than the best Pauli approximations suggest.