QEC Decoding

NVIDIA Ising Decoding

NVIDIA Ising Decoding is part of the NVIDIA Ising open model family. It provides AI-based pre-decoder models for quantum error correction (QEC) on surface codes.

The pre-decoder is a 3D convolutional neural network that consumes detector syndromes across space and time, predicts corrections that reduce syndrome density, and passes residual syndromes to a standard global decoder (such as PyMatching) for final correction. Two open models are available:

• Ising Decoder SurfaceCode 1 Fast (0.9M parameters, receptive field R=9) for low-latency decoding
• Ising Decoder SurfaceCode 1 Accurate (1.8M parameters, receptive field R=13) for higher accuracy

Both models are pre-trained with a depolarizing noise model for surface codes of any distance. Ising Decoding is available through the Control API for experimentation. Note that this hosted deployment is intended for testing and evaluation only, not for production QEC workflows. Real-time decoding requires sub-microsecond latencies constrained by qubit coherence times, which are only achievable with dedicated GPU hardware co-located with the quantum processor.

Open weights are also available on Hugging Face and GitHub.

Available models

Usage

The Ising Decoding models accept detector syndrome tensors as 5D NumPy arrays with shape (batch, 4, T, D, D) where T is the number of measurement rounds and D is the code distance.

1
import numpy as np
2
from conductorquantum import ConductorQuantum
3
4
client = ConductorQuantum(token="YOUR_TOKEN")
5
6
syndrome = np.zeros((1, 4, 3, 3, 3), dtype=np.float32)
7
8
result = client.control.models.run(
9
    model="ising-decoding-v1-fast",
10
    data=syndrome,
11
)

For higher accuracy at the cost of increased latency, use the accurate variant:

1
result = client.control.models.run(
2
    model="ising-decoding-v1-accurate",
3
    data=syndrome,
4
)

Key results

Integrated with uncorrelated PyMatching, the pipeline achieves end-to-end decoding runtimes on the order of O(1 us) per syndrome measurement round at large code distances on NVIDIA GB300 GPUs, while reducing logical error rates relative to global decoding alone. The larger model outperforms correlated PyMatching up to distance-13.

Key performance highlights from the model architecture paper:

• 2.5x improvement in speed and 3x improvement in accuracy over the prior state of the art
• Runtimes below 1 us per round are achievable with temporal parallel block-wise decoding across multiple GPUs
• Models trained at modest distance generalize to much larger distances, since the 3D CNN operates locally

Training framework

The Ising Decoding repository includes a complete training framework built on PyTorch and CUDA-Q, allowing developers to train or fine-tune pre-decoders for their own noise models and hardware. The framework supports ONNX export, INT8/FP8 quantization, and TensorRT deployment for production inference.