Nvidia opens Ising models

Quantum computers are powerful in theory, but today’s machines drift out of tune and make frequent mistakes. Nvidia said on April 14 it is open-sourcing a new Ising model family to help fix both problems. (nvidianews.nvidia.com) The release covers two jobs that sit between a lab demo and a usable machine: calibration and error correction. Nvidia said Ising Calibration can cut retuning work from days to hours, while Ising Decoding is aimed at real-time correction of qubit errors. (nvidianews.nvidia.com) Calibration is the constant tuning that keeps a quantum processor operating, like readjusting a musical instrument as the room changes. Error correction is the separate step that catches and interprets mistakes from noisy qubits before those mistakes overwhelm a calculation. (datacenterdynamics.com) Nvidia’s developer page describes Ising Calibration as a vision-language model that reads measurements from a quantum processor and can be used in an automated workflow. It describes Ising Decoding as a set of three-dimensional convolutional neural network models trained to decode surface-code errors with low latency. (developer.nvidia.com) The company said its decoder models are up to 2.5 times faster and up to 3 times more accurate than traditional approaches, and its developer page compares them with the open-source decoder pyMatching at a code distance of 13 and physical error rate of 0.003. Those claims come from Nvidia’s own benchmarks, not an independent head-to-head test published in the announcement. (nvidianews.nvidia.com) (developer.nvidia.com) The models are part of a bigger Nvidia push to make quantum systems look more like hybrid computers than standalone exotic hardware. Nvidia’s CUDA-Q platform already lets researchers combine central processing units, graphics processing units and quantum processing units in one workflow, and the company says Ising plugs into that stack. (developer.nvidia.com) (nvidia.com) That stack also includes Nvidia’s NVQLink interconnect and its Accelerated Quantum Computing Research Center, where the company is pairing partner quantum hardware with a GB200 NVL72 system. Nvidia says the goal is to link quantum processors to artificial intelligence supercomputers for low-latency control and error correction. (nvidia.com) Nvidia said early users of Ising Calibration include Atom Computing, IonQ, Infleqtion, IQM Quantum Computers, Lawrence Berkeley National Laboratory’s Advanced Quantum Testbed and the United Kingdom National Physical Laboratory. It said Ising Decoding has been deployed by Cornell University, Sandia National Laboratories, the University of Chicago, the University of Southern California and Yonsei University, among others. (datacenterdynamics.com) (nvidianews.nvidia.com) Nvidia is shipping more than model weights. Its Ising package includes a training framework, benchmark data, example workflows, and a “cookbook” for fine-tuning, quantization and deployment, with downloads linked through GitHub and Hugging Face on the company’s developer site. (developer.nvidia.com) The pitch is straightforward: quantum hardware still needs a lot of classical help. By open-sourcing tools for tuning and error repair, Nvidia is betting that one of the hardest parts of quantum computing will be managed with more conventional artificial intelligence software. (nvidianews.nvidia.com) (developer.nvidia.com)