Towards real time, low latency, circuit level decoding of QLDPC codes

We extend Astra, our graph neural network approach for decoding QLDPC codes, and showcase the scalability and low latency performance for circuit level depolarizing noise. Previous work [https://arxiv.org/abs/2408.07038] demonstrated that for code capacity depolarizing noise, Astra is outperforming the BP+OSD decoder on accuracy and speed. Moreover, without sacrificing decoding performance, our decoder can perform extrapolated decoding, i.e. it can decode high distances using models trained on lower distances. Herein, we present results for both surface codes and bivariate bicycle codes. We perform latency and hardware acceleration studies of our decoder, and illustrate a convincing path towards Astra’s capability for real time decoding.