NVIDIA’s Ising Models

Quantum computers use qubits, which are fragile bits that can drift out of tune or flip by mistake after tiny bursts of noise. NVIDIA said on April 14 it is releasing an open model family called Ising to help tune those machines and catch those errors faster. (nvidianews.nvidia.com) The release starts with two tools aimed at the two chores quantum teams repeat constantly: calibration and decoding. NVIDIA said Ising Calibration is a 35 billion parameter vision-language model, while Ising Decoding ships as two three-dimensional convolutional neural networks with 0.91 million and 1.79 million parameters. (nvidia.com) Calibration is the process of reading a quantum chip’s measurements and nudging its settings back into place, the way a technician retunes an instrument after it slips off pitch. NVIDIA said its calibration model reads experimental plots from quantum processing units and can cut calibration time from days to hours. (nvidianews.nvidia.com) Decoding is the real-time cleanup step after errors appear, where a classical computer has to infer what went wrong before more mistakes pile up. NVIDIA said its decoding models are up to 2.5 times faster and 3 times more accurate than traditional approaches such as pyMatching. (nvidia.com; siliconangle.com) Those tasks sit at the center of quantum computing’s biggest engineering problem: qubits are noisy enough that useful systems need constant tuning and layers of error correction. NVIDIA’s developer blog said leading processors still make about one error in every thousand operations, while useful machines would need error rates closer to one in a trillion. (developer.nvidia.com) NVIDIA is packaging the models as open tools rather than closed services. The company said the models are released with permissive licensing, documented data provenance, training methods, data sets, and tools for retraining, fine-tuning, quantization, and deployment. (nvidia.com; github.com) That matters for quantum hardware makers because each machine has its own noise patterns, control electronics, and proprietary measurement data. NVIDIA said users can adapt the base models to their own hardware while keeping quantum processor data on-site. (developer.nvidia.com) The company is also tying Ising into a larger push to link quantum processors with graphics processing unit supercomputers. NVIDIA’s product pages place Ising alongside CUDA-Q software and the NVQLink interconnect as part of what it calls a quantum-graphics processing unit supercomputing platform. (nvidia.com; nextplatform.com) Outside groups are already attached to the launch. NVIDIA named adopters including Academia Sinica, Fermi National Accelerator Laboratory, Harvard John A. Paulson School of Engineering and Applied Sciences, IQM Quantum Computers, Lawrence Berkeley National Laboratory’s Advanced Quantum Testbed, Infleqtion, and the United Kingdom’s National Physical Laboratory. (nvidianews.nvidia.com) One of the training and benchmark efforts came from Northwestern University and Fermilab, which said their superconducting qubit data helped train Ising Calibration. Northwestern said the benchmark tests the model on six tasks, including figure description, outcome classification, physical interpretation, data-quality assessment, charge-jump extraction, and stability judgments. (quantum.northwestern.edu) The pitch is straightforward: use artificial intelligence to handle the repetitive, high-speed control work that noisy qubits still cannot survive without. Whether Ising becomes a standard tool will depend on how widely quantum labs adopt and retrain it, but NVIDIA has now put that software into the open. (github.com; nvidianews.nvidia.com)