Performance Studies

In-depth performance studies of CUDA-Q QEC decoders on NVIDIA GPUs – measuring decode latency, logical error rate, and the trade-offs behind decoder tuning knobs.

• Gamma ensembles – how ensembling Relay BP gamma trajectories narrows the decode-latency tail, improving the logical error rate under hard decode deadlines by up to ~89x on bivariate-bicycle codes (measured on a single GB200 with CUDA-Q QEC 0.7.0).

• Relay solution recording – how recording every Relay BP convergence replaces a per-stop_nconv sweep of full decode runs with one recording run plus offline post-processing, reproducing every RelayBP-N result exactly.

• Min-LLR OSD initialization – how osd_init_method="min_llr" lets a 10-iteration BP+OSD decoder beat a 60-iteration one by up to ~100x in logical error rate on joint-XYZ circuit-level DEMs of bivariate-bicycle codes, and makes correlated (joint XYZ) BP+OSD decoding more accurate than uncorrelated (split X/Z) decoding.

• NV-Fusion latency – how overlapping decoding with syndrome acquisition bounds the final-round response latency of long-running surface-code experiments.

• Improving Relay BP Decoding With Gamma Ensembles
• Sweeping Relay BP Stopping Criteria From a Single Run
• Improving BP+OSD Decoding With Min-LLR OSD Initialization
• Bounding Real-Time Decode Latency With NV-Fusion