Nvidia slaps forehead: I know what quantum is missing – it's AI!
One error in every thousand operations is one too many
Quantum computers promise major speedups for problems in materials science, logistics, and financial modeling, but first they need to be made reliable, something Nvidia believes its AI models can help with. When you've got a GPU hammer, every problem starts to look like an AI nail.
On Tuesday, the GPU slinger unveiled its new open weights models aimed at helping quantum hardware developers drive down processor error rates.
According to Nvidia, even the best quantum systems generate errors roughly once in every thousand operations. To make them truly useful, they contend the error rates will need to come down by a factor of a billion.
The first of its new quantum models, codenamed Ising Calibration, does just what its name implies. The GPU giant says the 35 billion-parameter vision-language model was trained on data generated by partner systems, to help developers dial in the ideal settings to minimize noise within the systems.
Nvidia claims the model could be integrated into an agentic framework to fully automate this process by streaming data collected by the system and making adjustments until error rates fall below certain thresholds. In this respect, it's a bit like quantum autotune.
Unlike many large language models, Ising Calibration is fairly lightweight and can easily be run on an RTX Pro 6000 Blackwell or an Nvidia GB10-based system like the DGX Spark.
While Ising Calibration can help reduce how often errors occur, it can't eliminate them entirely. This is where Nvidia's Ising Decoding models come in. They are available in two sizes, which once trained, are designed to detect and correct errors in real time.
To make this possible, Nvidia employed an older convolutional neural network (CNN) architecture. Compared to Ising Calibration, these models are tiny, coming in at 912,000 parameters for Ising-Decoder-SurfaceCode-1 and 1.79 million for the larger "Accurate" model, allowing them to catch errors between 2.25 and 2.5x faster than conventional approaches using frameworks like PyMatching.
Weights for Ising Calibration 1 and Ising Decoder SurfaceCode 1 are on Hugging Face, with Ising Calibration 1 also landing on Nvidia Build and as an inference microservice (NIM). Alongside the models, Nvidia is also rolling out training frameworks to help developers generate synthetic data and fine tune the models for their specific systems, and inference blueprints for implementing the models.
The models are only the latest in a slew of investments Nvidia has made in quantum computing over the past few years, which include everything from hardware and software libraries to a research center with a Blackwell-based supercomputing cluster. ®