Quantum computing has the potential to solve problems beyond the reach of today's computers, from drug discovery to materials science and optimisation. But before that can happen, the industry must first overcome the errors that are created by quantum computing’s extreme sensitivity to noise.
Quantum error correction (QEC) detects and corrects these errors. It’s essential for building reliable and useful quantum computers, but this also creates enormous demands for real-time data processing. Future systems may need to process terabytes of information every second, with near-perfect accuracy, to support error-free quantum computing.
To explore how to meet these demands, Riverlane and Nvidia are collaborating on the use of AI and NVIDIA accelerated computing for QEC, including how GPUs can make decoder calibration faster, and help quantum systems keep up with changes in hardware behaviour.
Keeping up with changing quantum hardware
A decoder is the part of a QEC system that interprets the signals produced by a quantum processor. It works out which errors are most likely to have occurred and tells the system how to correct them.
To do so accurately, the decoder needs a reliable picture of the errors that occur. It uses numerical settings, known as weights, to estimate how likely different types of errors are and decide which corrections to apply. These settings are based on measurements from the quantum hardware, so they remain accurate only while the qubits behave as they did when the measurements were taken.
The challenge is that quantum hardware is far from perfectly stable. Qubit frequencies shift, defects can appear or disappear, and gate performance can change over time. As the noise changes, the decoder’s settings can gradually become less accurate, reducing the effectiveness of error correction.
Because of this instability, regular calibration is essential. This can be slow, and every minute spent calibrating is time when the quantum computer is not running an experiment or doing useful work.
Building a faster and more efficient calibration loop
Riverlane has developed an AI-based calibration loop that learns improved decoder settings directly from experimental data. The system can update the decoder’s understanding of the hardware without relying on manual tuning. As part of a collaboration, Riverlane uses NVIDIA H100 GPU to test how well different decoder settings perform.
The calibration process starts with an informed estimate of the likely error patterns. The AI system then proposes different decoder settings, tests them against data from the quantum hardware and favours the settings that lead to fewer logical errors. Over time, the decoder’s picture of the hardware becomes more accurate. The resulting improvement comes from the experimental data itself, without the need for manual adjustment of individual settings.
The results show the potential for calibration to become faster and more routine. Using Riverlane’s Local Clustering Decoder (LCD), adapted for high-throughput GPU processing, a training cycle on the surface code that previously took three minutes using a single-core CPU now takes just 2.8 seconds.
The H100 GPU can decode more than four million surface code syndromes per second, compared with around 59,000 per second on a single CPU core. This makes the GPU particularly well suited to calibration, with the ability to process large batches of independent measurements in parallel.
There are multiple benefits to faster calibration. Speed changes how often calibration can happen, for starters. A loop that takes seconds rather than minutes can run before every experiment, rather than once a day. It can also train for longer and produce better weights within the same timeframe.
For manufacturers of quantum computers, that means less idle time and fewer drops in logical performance. For HPC centres and control-system companies, it’s a concrete, well-defined workload for the GPUs they are already placing beside quantum hardware.
Using GPUs to create the quantum-classical systems of the future
GPUs are already widely used across industries, and high-performance computing (HPC) centres have deployed them at scale. They are also beginning to play a role in the quantum control stack, where fast processing is essential for managing increasingly demanding quantum systems.
As QEC experiments run for longer and use larger numbers of qubits, changes in hardware behaviour become increasingly important. The decoder needs to keep up with this changing environment while meeting the strict speed and latency requirements of real-time error correction.
This investigation also covers broader questions about the future of quantum-classical computing. Riverlane will evaluate AI-enhanced approaches to error correction, explore how different computing technologies can support future quantum processors, and investigate how Riverlane’s QEC technologies could work alongside NVIDIA accelerated computing and quantum platforms.
“Quantum error correction is the key technology that will enable useful quantum computers,” Steve Brierley, founder and CEO of Riverlane, adds. “As quantum systems scale, QEC becomes one of the world’s most demanding real-time computing challenges. Through this collaboration with NVIDIA, we are exploring how advances in quantum computing, AI and accelerated computing can come together to support the next generation of quantum systems.”
“Accelerated computing is becoming the engine of useful quantum computing, transforming how quantum computers are built and operated,” said Sam Stanwyck, Director of Quantum Product at NVIDIA. “Riverlane is using AI to help error correction keep pace with changing quantum hardware. Accelerating that learning process opens the door to more frequent calibration, so researchers can spend less time tuning their quantum computers and more time doing science.”
Recalibrating a decoder is essentially a learning problem, and a GPU makes that learning fast enough to be routine.
Riverlane will continue to investigate how NVIDIA accelerated computing can help make quantum error correction faster, more adaptive and more practical for utility-scale quantum computers. Ultimately, this paves the way for quantum computers to achieve their potential.