Alice & Bob decodes LDPC-cat error correction codes on GPUs via NVIDIA NVQLink

[Seb93]

French cat qubit company Alice & Bob says it is the first to decode its LDPC-cat error correction codes on GPUs using NVIDIA's NVQLink architecture. Quantum Zeitgeist covered the announcement, which is based on a press release. Decoding is the classical work of reading error signals from a quantum chip and working out how to fix them, and it has to happen very quickly

The team used NVQLink to connect a Quantum Machines OPX1000 controller and its Open Acceleration Stack to GPUs. The round trip latency was in the microsecond range, which is the kind of speed real time error correction needs. The goal was to test whether GPU decoding is a practical option for this type of code

Some background helps here. LDPC-cat codes were set out in a January 2025 paper about low overhead quantum computing in two dimensions. They combine cat qubits, which naturally resist one type of error, with efficient LDPC codes to handle the rest. In March 2026, Alice & Bob decoded simulated syndrome data on NVIDIA's CUDA-Q in just under two hours

The new result is a proof of concept, and the company is clear about its limits. It still uses synthetic data rather than signals from a real chip, and it does not yet run in real time. The team is now speeding up the pipeline that moves measurement data to the GPU and believes processing syndromes in parallel could help close the gap

Statements came from Sam Stanwyck of NVIDIA and Juliette Peyronnet of Alice & Bob. NVIDIA has been building partnerships across the quantum industry, and decoding is a natural fit for GPUs. Each step like this makes fault tolerant machines feel a little more concrete

Do you think GPUs will become the standard way to decode quantum error correction? And how close are cat qubits to a real fault tolerant demonstration?

Alice & Bob Decodes LDPC-cat Codes On GPUs With NVIDIA NVQLinkquantumzeitgeist.com

[Vieira76]

Will FPGAs and custom chips beat GPUs for this in the end? Some groups think dedicated decoders will be needed for the fastest qubits

Cat qubits are relatively slow, so GPUs may be enough for them. Different hardware may need different answers. Flexibility counts for something too. GPUs can run new decoders with a software update

[Pixel Mark]

Paris has become a real hub for quantum hardware. Alice & Bob, Pasqal and Quandela all in one city is impressive. Good for European competition. French government support has helped a lot. The national quantum strategy was well funded early on

[BookerT]

NVIDIA is turning up in almost every quantum announcement now. They have positioned GPUs as the classical half of every hybrid system. Smart strategy, hard to argue with

[Cobalt Sophie]

Synthetic data only, so let's not get carried away. Real hardware syndromes will be noisier and less predictable. Still, you have to build the pipeline before you can test it on a live chip. Each piece has to work before the whole thing can

[Sarah87]

Cat qubits are one of the most elegant ideas in the field. Suppressing bit flips in hardware means the error correction only has to handle phase flips. That should cut the overhead dramatically. Alice & Bob's roadmap depends on that advantage holding up at scale

So far, the signs are good. Bit flip times keep getting longer in their published results

[CoastalProtocol]

Microsecond latency over NVQLink is the number that matters. Decoding is useless if it cannot keep up with the chip

Getting into that range with off the shelf GPUs is encouraging. Cat qubits are slower than transmons, which gives a little more breathing room. That helps here

[CodyRhodes29]

Going from two hours of offline decoding in March to a near real time setup in October is quick progress. The next milestone is obvious. Run it on real syndromes from their own chips. Momentum matters for a startup