IonQ Demonstrates Real-Time Quantum Error Decoding on a Single CPU

Decoding remains a bottleneck for scalable quantum computing

Researchers are advancing system architectures, software, algorithms, and workflows for classical-quantum hybrid environments. However, error detection and correction remain major obstacles to fault-tolerant, scalable quantum computing. Without effective error correction, quantum systems produce too many errors to support practical use.

Logical qubits are central to this effort. They combine many noisy physical qubits and provide more reliable building blocks for quantum systems. Another essential component is decoding, the process of analyzing error signals, known as syndromes, to identify the errors affecting a system.

Quantum computers cannot measure qubits directly without destroying their state. Qubits are highly sensitive to environmental noise, including light, vibration, and interference from other qubits. Such disturbances can eliminate the quantum state and the data it contains. Instead, systems measure related data points to produce syndromes that indicate whether an error has occurred. Decoders then analyze those syndromes.

According to scientists with QuEra, which uses neutral atoms to build qubits, decoding must manage increasing computational complexity as quantum systems grow. Decoders need to balance speed and accuracy so they can operate in real time while interpreting complex error patterns correctly.

Decoding relies heavily on classical computing algorithms. It has historically required significant processing power, and the demand increases with the number of qubits. Decoders must also be designed for the specific error-correction code in use, such as surface codes or color codes.

GPUs and AI support quantum error correction

Several approaches are being explored to improve decoding performance. GPU acceleration, including Nvidia’s CUDA-Q platform, is one option. Artificial intelligence is another.

IBM recently used a large language model framework to evaluate thousands of code variations and identify suitable implementations for particular environments. Google created AlphaQubit, which uses a transformer-based neural network to decode and correct errors. Google has also described an AI-based reinforcement-learning technique that processes encoded information collected by a quantum computer, learns from detected errors, and adjusts the system accordingly.

That approach could allow superconducting quantum computers to run for longer periods without stopping for calibration adjustments whenever errors are detected.

IonQ’s single-CPU decoder

Scientists at IonQ, which is developing trapped-ion quantum systems, said they created a decoder capable of handling the process in real time on a single standard CPU. The decoder can operate in the background, allowing the quantum computer to continue running.

In a research paper published on arXiv, the scientists described the goal of a MegaQuOp quantum computer as a system capable of executing millions of operations across hundreds of logical qubits. They identified real-time decoding for hundreds of logical qubits during computations involving millions of logical operations as a major challenge.

IonQ executives said the work indicates that classical hardware may not need to scale exponentially as quantum computers add logical qubits. They described single-CPU decoding as an important part of the company’s roadmap, which extends beyond the 256 physical qubits in its Superion system, introduced earlier this month.

The company also announced that Florida International University is buying a Superion system. The system is scheduled for installation in late 2027 at a specialized facility on the university’s campus.

Benchmark results

For the research, the scientists created benchmark circuits simulating up to 408 logical qubits across 88 memory blocks and magic factories. Magic factories are subsystems in fault-tolerant quantum systems that create the conditions needed for universal quantum computing.

The circuits executed more than 31.5 million individual quantum operations at the MegaQuOp scale while adding as little as 0.02 percent in stretch time. The researchers said this represented very little additional delay for the overall computation.

Other research has focused on accelerating surface-code and quantum low-density parity-check (LDPC) memory decoders through software and specialized hardware. These efforts use GPUs, application-specific integrated circuits, field-programmable gate arrays, and tensor processing units to achieve the latency required by superconducting quantum systems.

IonQ’s decoding system covered 68 LDPC memory blocks and 20 magic state factories created with 11,680 physical qubits. It uses a dual sliding-window decoder based on a beam search decoder. This error-correction algorithm decodes LDPC codes by retaining a fixed number of the most promising error paths, known as the beam width, at each step.

The system generates detector error models (DEMs) in real time and uses memory optimizations that allow multiple decoder instances to run concurrently on one CPU. DEMs are mathematical descriptions of how gate errors or noise produce particular error detectors in an error-correction circuit.

The Walking Cat architecture

A key part of the CPU-only approach is IonQ’s Walking Cat architecture for trapped-ion quantum systems, which the company introduced in a blueprint in April. The architecture includes a compiler, error-correction protocols, a microarchitecture, complete simulations, and a fast decoder, all based on LDPC error-correction codes.

The architecture centers on a “cat factory,” which creates cat states. These states are superpositions of distinct states that exist simultaneously and can be distributed throughout the system to perform logical operations.

The researchers emphasized the architecture’s comparatively simple structure. Logical computation does not require merging or deforming qLDPC memory blocks or modifying their syndrome-extraction circuits.

Instead, cat-based measurements are inserted into a regular stream of syndrome extraction. A large class of Clifford operations, which are quantum gates that map Pauli operators, X, Y, and Z, to other Pauli operators under conjugation, is handled through software frame tracking.

According to the researchers, this design makes the full classical decoding pipeline, from online DEM generation through decoding, feasible at MegaQuOp scale using only 12 cores of a single commodity CPU.