The Signal
NVIDIA releases QEC decoder, claims up to 13,000× speedup
NVIDIA released the TCBP-JOBS quantum error-correction decoder through its CUDA-Q QEC 0.9.0 library. The company claims up to 13,000× faster decoding in a specific comparison with another decoder, while its research paper reports lower speedups against other settings.

A speedup with a specific benchmark
NVIDIA released the TCBP-JOBS quantum error-correction decoder through its CUDA-Q QEC 0.9.0 library, claiming it can accelerate some decoding research by up to 13,000×. The company’s Nov. 16 announcement says the algorithm could cut a computation for evaluating quantum-computing designs from 12 weeks to 45 minutes.
The larger speedup is not a general result across every decoder or workload. In the research paper, NVIDIA’s authors report that the 13,000× figure came from comparing TCBP-JOBS with an unsparsified, long-beam configuration of the Tesseract decoder on bivariate-bicycle codes; against Tesseract’s sparsified presets, the reported speedups range from 10× to 7,400×.
How the decoder works
TCBP-JOBS combines two methods. Trajectory-calibrated belief propagation, or TCBP, learns calibrated probabilities for possible faults from the paths taken during belief-propagation calculations. Jittered OSD beam search, or JOBS, uses those probabilities to guide multiple searches in parallel on GPUs.
A quantum error-correction decoder processes measurement data to infer likely errors and identify corrections. NVIDIA’s researchers tested their approach on memory and surgery circuits using surface, color and bivariate-bicycle codes, and compared its decoding accuracy with Tesseract, a most-likely-error decoder.
Across the tested code settings, the paper reports logical error rates comparable to Tesseract’s and, for the largest codes, rates up to 4.5× lower. Those results have a stated operating range: the comparison describes physical error rates from 0.1% to 0.4%. At lower physical error rates, the authors report that Tesseract was more accurate in their rare-event-splitting tests on the gross code.
Benchmark workload and limits
One reported workload involved decoding 24.2 million shots from a Tesseract benchmark. NVIDIA’s authors say TCBP-JOBS completed it in 45 minutes on one GB200 GPU, while the Tesseract run required 130,257 CPU thread-hours, which the paper equates to about 12 weeks on a 64-core server.
The paper’s speed and accuracy results come from research workloads, including simulated circuits and comparisons with specified Tesseract configurations. The release frames faster research decoding as a way to evaluate candidate error-correction approaches and compare codes and decoder settings; it does not describe the 13,000× figure as a general speedup for operating quantum processors.
Part of a broader collaboration
The decoder release was included in NVIDIA’s announcement of planned collaborations with IBM and the U.S. Department of Energy’s Fermi National Accelerator Laboratory. The partners are developing a modular quantum-computing codesign platform intended to support work on error-correction codes, decoding algorithms and quantum-classical applications.
NVIDIA said DGX B200 systems are being colocated with IBM quantum computers, including systems powered by IBM Quantum Nighthawk, as part of the work. The platform is also intended to help researchers study how classical computing systems can scale alongside multiple linked quantum processors.