NVIDIA's Ising Decoder Slashes Color Code Logical Error Rates 300x+

NVIDIA has released Ising Decoder ColorCode 1 Fast, a 3D CNN-based pre-decoder built for triangular color codes. At code distance d=31 and a 0.3% physical error rate, it delivers over 347.7x better logical error rate (LER) and 7.3x faster runtime than Chromobius, the current open-source reference decoder for color codes.

Color codes can outperform surface codes for fault-tolerant logic thanks to transversal Clifford gates and simpler lattice surgery, but their decoding complexity has kept them sidelined. NVIDIA's training pipeline—built on cuQuantum, cuStabilizer, and PyTorch—trains a ~2.9M-parameter, 17-layer model on synthetic syndrome data, offering a practical path around that bottleneck.

Model weights are published on Hugging Face under the OpenMDW 1.1 license, with the full training architecture open-sourced on GitHub under Apache 2.0. The release gives QPU builders a way to train custom decoders tuned to their own noise characteristics, potentially reviving color codes as a viable fault-tolerant QEC scheme.