Most GPU-decoder coverage on this site so far has been about NVIDIA's own claims for its own tools. This is a different kind of story: independent companies running their own error-correction workloads on NVIDIA's CUDA-Q QEC library and publishing their own numbers, which carries more weight than a vendor benchmarking its own product.
Alice & Bob: 18 hours to under 2
In March 2026, Alice & Bob reported a 9.25x speedup decoding quantum error correction data by moving the workload onto GPUs through CUDA-Q, cutting a run that took 18 hours on CPU down to under 2 hours. The company reported the GPU-accelerated version produced the same logical error performance as the CPU baseline, meaning the speedup came from throughput, not from a shortcut that traded accuracy for speed. That distinction matters: a decoder that's fast because it's cutting corners isn't solving the real-time decoding problem this site has covered as the real bottleneck standing between today's hardware and fault tolerance.
Quantum X Labs: a transformer decoder against the classical baseline
In July 2026, Quantum X Labs ran its Deep Quantum Error Correction workflow on GPU in an AWS environment, benchmarking a transformer-based decoder called QECCT against minimum-weight perfect matching, the classical decoding baseline this site already tracks as the maintained standard. QECCT reportedly outperformed MWPM in selected simulated toric-code noise configurations, worth reading precisely: "outperformed in selected configurations" is a real, specific, checkable claim, and a narrower one than "outperforms MWPM," which the coverage doesn't say.
What CUDA-Q QEC 0.5.0 added
The underlying library both companies built on picked up real capability in its 0.5.0 release: online real-time decoding, GPU-accelerated algorithmic decoders, higher-performance AI decoder inference infrastructure, sliding-window decoder support, and RelayBP, an improvement on BP+OSD decoding for qLDPC codes that adds memory-strength damping per graph node to help convergence. That's the same qLDPC code family our coverage of USTC and Origin Quantum's routing codes discussed from the code-design side. A better decoder and a better code are separate contributions to the same overhead problem, and both are needed together for either to pay off in practice.
Why third-party adoption is the more interesting story here
A vendor's own benchmark tells you what a vendor wants you to know. Two independent companies choosing to build their production decoding workflows on the same GPU infrastructure, and being willing to publish their own numbers rather than only NVIDIA's, is a different and stronger kind of evidence. It doesn't make either company's specific number independently verified in the peer-review sense, but it does mean the underlying tool is getting real use outside NVIDIA's own marketing, which is the actual test for whether infrastructure like this sticks.
What to watch next
Whether more companies publish their own before-and-after numbers using CUDA-Q QEC's decoders, and whether any of these results eventually go through formal peer review rather than only a press release or blog post. Our picks-and-shovels supply chain piece covers NVIDIA's broader infrastructure strategy, of which this decoder ecosystem is one concrete, checkable piece.