Riverlane and NVIDIA use H100 GPUs to speed up quantum error-correction decoder recalibration 65-fold

Riverlane/Deltaflow

Riverlane and NVIDIA reported benchmarks showing GPUs can recalibrate quantum error-correction decoders much faster. Decoder weights track hardware error rates, which drift over time, so they must be recalibrated between experimental runs; the GPU loop learns better settings from experimental data.

• Training epoch time fell from about 180 seconds on a single CPU core to 2.8 seconds on one NVIDIA H100, a 65x speedup for the surface code.
• Syndrome throughput rose from about 59,000 per second on one CPU core to more than 4,200,000 per second on the H100, a 71x gain.
• Calibrating against device drift cut logical error rates by several percent up to a factor of two versus uncalibrated priors, depending on the code.
• The page presents GPUs as the offline learning layer, with real-time decoders such as Riverlane's Deltaflow still handling decoding during circuit execution; it gives no timelines.