How to choose a decoder for a fault-tolerant quantum computer? The speed vs accuracy trade-off

Achieving practical quantum advantage requires a classical decoding algorithm to identify and correct faults during computation. This classical decoding algorithm must deliver both accuracy and speed, but in what combination? When is a decoder "fast enough" or "accurate enough"? In the case of surface codes, tens of decoding algorithms have been proposed, with different accuracies and speeds. However, it has been unclear how to choose the best decoder for a given quantum architecture. Should a faster decoder be used at the price of reduced accuracy? Or should a decoder sacrifice accuracy to fit within a given time constraint? If a decoder is too slow, it may be stopped upon reaching a time bound, at the price of some time-out failures and an increased failure rate. What then is the optimal stopping time of the decoder? By analyzing the speed vs. accuracy tradeoff, we propose strategies to select the optimal stopping time for a decoder for different tasks. We design a protocol to select the decoder that minimizes the spacetime cost per logical gate, for logical computation of a given depth. Our protocol enables comparison of different decoders, and the selection of an appropriate decoder for a given fault-tolerant quantum computing architecture. We illustrate our protocol for the surface code equipped with a desktop implementation of the PyMatching decoder. We estimate PyMatching is fast enough to implement thousands of logical gates with a better accuracy than physical qubits. However, we find it is not sufficiently fast to reach 10^5 logical gates, under certain assumptions, due to the decoding delay which forces qubits to idle and accumulate errors while idling. We expect further improvements to PyMatching are possible by running it on a better machine or by reducing the OS interference.