Fault-tolerant logical operations are required for useful quantum computers, so researchers study many quantum error-correcting (QEC) codes to reduce Logical Error Rates (LER) on quantum processing units (QPUs). Among topological codes, surface codes are well understood for lattice-surgery-based logical operations, but they are not always optimal for fault-tolerant logical computation. Color Codes are another topological-family option: they typically need more physical qubits for the same memory-target LER than surface codes, but can implement logical gates more efficiently because all Clifford gates can be applied transversally and their symmetry simplifies lattice-surgery operations.
Historically, color codes have been less used because decoding them is harder than decoding surface codes, and fast, accurate, real-time decoders were not readily available. NVIDIA introduced "Ising Decoder ColorCode 1 Fast," a 3D convolutional neural network (CNN) pre-decoder targeted at triangular color codes, intended to speed up and improve full decoding pipelines.
Claimed improvements and use case
NVIDIA reports that, when used as a pre-decoder in front of the open-source Chromobius decoder, Ising Decoder ColorCode 1 Fast can deliver substantial gains: for code distance d = 31 and a physical error rate of p = 0.3%, they report more than 347.7× improvement in LER and a 7.3× faster runtime compared with Chromobius alone. The pre-decoder sparsifies localized space–time error syndromes and predicts full space–time corrections locally, which enables blockwise, parallel decoding architectures suited for real-time error correction during lattice-surgery and algorithm execution.
Training pipeline and example
The training pipeline uses NVIDIA cuQuantum together with the cuStabilizer library and PyTorch to generate synthetic training data on the fly and to train 3D CNNs tuned to a user’s noise model and code distance. Users specify the noise model, triangular color code distance, and model depth (deeper models generally increase accuracy but at higher inference cost). The architecture supports trading off the number of layers for runtime versus accuracy based on the code distance, physical error rate, global decoder effectiveness, and latency budget.
NVIDIA provides a short example: building a distance-5 color-code memory experiment with superdense syndrome extraction and Si1000 circuit-level noise at p = 1e-3 (0.1%). In that example, decoding with Chromobius yields an LER around 3.6e-03 (≈7.3e-04 per round). From such a baseline one can sweep distances and error rates to obtain threshold curves, or insert the Ising-Decoder-ColorCode-1 pre-decoder before Chromobius to sparsify syndromes prior to final decoding.
Model details
The ColorCode 1 Fast model has the following characteristics: 17 layers, a receptive field of 13, trained on input volumes of size 13×13×19 (although arbitrary input volumes can be used at inference time), and roughly 2,900,000 parameters. Its relatively small size makes it efficient to run on GPUs, but it delivers less LER improvement than larger models; therefore deployment requires selecting the right model size for the desired latency and accuracy trade-off.
Open resources and adoption
The NVIDIA Ising model family is released as open resources: weights, training architectures, synthetic data generation code, benchmarks, and recipes are provided so others can adapt, train, and fine-tune models for their specific QPU noise characteristics. The full training pipeline for Ising Decoder ColorCode 1 Fast is available under Apache 2.0, alongside a Color Code Model Architecture paper and additional analyses.
Why this matters
Fast, accurate decoders are essential to make alternative topological codes—like color codes—practical for fault-tolerant quantum computation. If the reported improvements generalize across relevant parameter regimes, color codes could become a competitive choice for logical operations alongside or instead of surface codes. The combination of open-source models and GPU-accelerated toolchains (cuQuantum, cuStabilizer, PyTorch) allows QPU builders and decoder developers to produce tailored pre-decoders that scale with their hardware and support real-time decoding during lattice-surgery and algorithm execution.
