Riverlane and NVIDIA speed up error correction decoder calibration on GPUs

Started by Rory_39, Oct 21, 2026, 12:49 AM

Previous topic - Next topic

0 Members and 1 Guest are viewing this topic.

Topic: Riverlane and NVIDIA speed up error correction decoder calibration on GPUs   Views(Read 64 times)

Rory_39

Riverlane and NVIDIA have published benchmarks showing how GPUs can dramatically speed up the calibration of quantum error correction decoders. Quantum Computing Report covered the announcement. It addresses a problem that will matter more and more as quantum computers grow

Decoders are the classical software that reads error signals from a quantum chip and works out how to fix them. They rely on weights that describe how likely each type of hardware error is. As qubits drift over time, those weights go out of date, and the decoder gets worse. Recalibrating them quickly is essential for keeping logical error rates low

The teams ran AI driven calibration loops on NVIDIA GPUs, using a GPU version of Riverlane's Local Clustering Decoder and building on earlier work from Google Quantum AI. For the surface code, one training epoch took about 180 seconds on a single CPU core and just 2.8 seconds on an NVIDIA H100, about 65 times faster. Throughput rose from around 59,000 to more than 4.2 million syndromes per second

Repetition code training ran 43 times faster, at about 155 million syndromes per second. One H100 was still three times faster than a 32 core CPU. Most importantly, recalibrating against device drift lowered logical error rates by several percent, and by up to a factor of two depending on the code

The picture that emerges is a sensible split of work. GPUs handle the heavy offline learning, while Riverlane's Deltaflow system handles real time decoding. We covered Alice & Bob's GPU decoding work earlier in the week, and it is clear that GPUs are becoming a core part of the quantum error correction stack

Are GPUs becoming as important to quantum computing as the qubits themselves? And how often will future machines need to recalibrate their decoders?


veritas.io

Halving logical error rates just by recalibrating the decoder is a big win. No new hardware needed, just better use of the data

That kind of improvement is cheap compared with building better qubits. Software gains like this compound with hardware gains
Coffee first. Questions later.

CollapseState47

Riverlane keeps positioning itself as the error correction layer for everyone. Working with NVIDIA strengthens that

Both companies want to be part of every quantum stack. Deltaflow is already in several labs. This adds a strong offline story on top

RVD17

Google's prior optimisation work being reused here is nice to see. Open research building on open research. That is how the field moves faster. Shared methods mean fewer wasted efforts across the industry

RayOfLight87

Sixty five times faster on one H100 is impressive, but how big are the codes? Larger codes will need much more compute. Hopefully the scaling stays favourable. Clusters of GPUs might be needed for big machines. Data centre GPUs are not cheap either

Taker92

Splitting offline learning onto GPUs and real time decoding onto dedicated hardware makes sense. Each part runs where it fits best. It is a good systems design. Real time and offline work have very different requirements. Mixing them would be a mistake

CharlotteFlair

Two GPU decoding stories in one week. NVIDIA has clearly decided error correction is a key market

Smart move, because every fault tolerant machine will need this. CUDA-Q is becoming the default glue for a lot of these projects. NVIDIA is building a moat in quantum software as well as hardware

RogueAI32

Drift is one of the most annoying problems in real quantum hardware. Qubits change from hour to hour. Fast recalibration means the decoder can keep up with those changes

Temperature changes, cosmic rays and two level defects all shift the error picture. Continuous calibration might become standard