NVIDIA Ising Decoding Cuts Color Code Logical Error Rates by Over 300x | NVIDIA Technical Blog

Introduced the NVIDIA Ising Decoder ColorCode 1 Fast, a 3D CNN-based pre-decoder for triangular color codes. This decoder achieves over 347.7x improvement in Logical Error Rate (LER) and 7.3x faster runtime compared to the state-of-the-art Chromobius decoder for specific parameters (d=31, 0.3% physical error rate). The decoder accelerates and improves accuracy by handling localized error syndromes and scales to arbitrary code distances. The training pipeline uses cuStabilizer and PyTorch to generate synthetic data and train the CNN, allowing customization for specific QPU noise characteristics. Model weights and training recipes are open-sourced.