Nvidia has released the code and training tools for the Ising Decoder ColorCode 1 Fast. This AI module preprocesses error signals before passing them to the Chromobius decoder, the company reports.
In simulations, this combination reduced logical error rates by 347.7 times and sped up processing by 7.3 times compared to using Chromobius alone. These results were achieved on a quantum memory model with a code distance of 31 and a physical error rate of 0.3%. The tests were conducted using synthetic data rather than on an operational quantum processor.
The Ising Decoder ColorCode 1 Fast is a 17-layer three-dimensional convolutional neural network with approximately 2.9 million parameters. Its receptive field is 13, and it was trained using input arrays sized 13 × 13 × 19.
This model does not function as a standalone decoder; instead, it serves as a preprocessing step. It analyzes local error signals, reduces their quantity, and passes the remaining sparse map to the classical Chromobius decoder.
Color Codes Move Closer to Real-Time Operation
Quantum error correction enables the combination of many unstable physical qubits into more reliable logical qubits. The decoder analyzes the results of control measurements to determine which errors need correction.
Surface codes are often used for storing quantum information due to their relatively high error threshold and simpler decoding process.
Color codes allow for more efficient execution of certain logical operations, but their error signals are more complex to process. According to Nvidia, the lack of fast and accurate decoders has long hindered the real-time application of such codes.
The Ising Decoder ColorCode 1 Fast is expected to reduce the load on the main algorithm. The authors of the study stated that the advantages of the combination with Chromobius increase as the code distance grows.
It's important to note that the speed comparison was conducted on different types of hardware. The neural network was run on the Nvidia DGX GB300, while Chromobius was executed on a Grace Neoverse-V2 processor. Therefore, the 7.3 times acceleration reflects not only differences between the algorithms but also the use of GPU instead of CPU.
Nvidia has posted the framework and training recipes in an open repository under the Apache 2.0 license.
In April, the company introduced a family of open Ising models, which includes tools for calibrating quantum processors and error correction.
In June, IBM unveiled an updated roadmap, stating that the company aims to create the world's first large-scale fault-tolerant quantum computer by 2029.