Quantum X Labs Tests AI Error-Correction Decoder on NVIDIA GPUs

Quantum X Labs says its AI-based error-correction decoder beat the standard classical one in tests, though only in simulation. The Nasdaq-listed company ran a transformer model it calls QECCT on NVIDIA GPUs, using NVIDIA’s CUDA-Q software for quantum error correction. The results are an early validation step, with no quantum hardware involved yet.

How the AI decoder compares with classical decoding

A decoder is the piece of software that keeps a quantum computer honest. Error correction spits out a stream of signals that flag when something has gone wrong, and the decoder reads those signals and works out the correction to apply, fast, before more errors accumulate. The trusted tool for that job is a classical algorithm called minimum-weight perfect matching, or MWPM. According to the July 24 press release, QXL’s bet, shared across parts of the industry, is that a trained neural network can do the same job as well or better, especially as error-correcting codes grow larger and the decoding gets harder.

QXL’s system, which it calls Deep Quantum Error Correction, uses a transformer, the same kind of model behind large language tools, trained on a code’s structure and its error signals to predict corrections. The company tested this QECCT decoder against MWPM in two simulated settings, a toric code and a surface code modeled on Google’s public geometry, run across several code distances. QECCT came out ahead of MWPM in selected conditions and held stable error rates under varying noise in others. That qualifier is worth keeping. It beat the classical decoder in some chosen regimes, and the company didn’t claim a clean sweep.

Why the results are still an early milestone

Everything here happened in software. The company frames it as an initial step and plans to move next to public experimental datasets, then to real syndrome data from superconducting processors through a collaboration with IQCC, a Quantum Machines company. It’s also weighing whether to run QECCT as a pre-decoder that clears easy errors before a slower decoder finishes, an approach similar to one NVIDIA has pushed with its own tools.

The claims are QXL’s own, without outside benchmarking, and the hardware phase is planned rather than done. Chief Quantum Technology Scientist Nir Sharon described the work as moving “from offline benchmarking toward practical real-time QEC.” The framing is fair enough. This is a staged plan that has cleared its first, simulated step, with the harder ones ahead.