Nvidia shares quantum tools

Quantum computing machines use qubits, which are fragile bits that can lose information from tiny disturbances like heat or vibration. Nvidia said on April 14 it released an open-source family of artificial intelligence models, called Ising, to help labs tune and correct those machines. (nvidianews.nvidia.com) The first models target two jobs that slow quantum labs down: calibration, which is repeated tuning of a processor, and decoding, which is the rapid interpretation of errors while a machine runs. Nvidia said Ising Calibration can cut tuning time from days to hours, and Ising Decoding is up to 2.5 times faster and 3 times more accurate than traditional approaches. (nvidianews.nvidia.com) Nvidia posted the models with pre-trained weights, datasets, training tools and instructions for retraining, fine-tuning and deployment. The company said the package is released under permissive licensing with documented data provenance and training methods. (nvidia.com) The field has a basic engineering problem: today’s qubits make mistakes often enough that useful systems need constant tuning and fast correction. Nvidia’s technical blog said leading processors still make an error about once in every thousand operations, while useful machines would need error rates closer to one in a trillion. (developer.nvidia.com) That is why Nvidia is pushing a hybrid design in which conventional chips do the cleanup work around the quantum chip. On its quantum platform page, the company says useful systems will combine quantum processing units with graphics processing units and central processing units rather than run on quantum hardware alone. (developer.nvidia.com) The calibration model is a vision-language model, meaning it reads charts and measurements the way a lab researcher would read plots on a screen. Nvidia said its first Ising Calibration model has 35 billion parameters and is tuned to infer calibration actions from experimental data from quantum processing units. (nvidia.com) Nvidia also published a benchmark called QCalEval to measure how well models understand those calibration plots. The company said the dataset includes 243 samples, 87 scenario types and 22 experiment families across superconducting and neutral-atom systems, and that its Ising Calibration 1 model reached a 74.7 zero-shot average score. (research.nvidia.com) For error correction, Nvidia is releasing two three-dimensional convolutional neural network models with 0.9 million and 1.8 million parameters. The company said they are designed for low-latency “pre-decoding” and ship with a training framework that can be adapted to different noise models through PyTorch and CUDA-Q. (nvidia.com) Nvidia said early adopters include Academia Sinica, Fermi National Accelerator Laboratory, Harvard John A. Paulson School of Engineering and Applied Sciences, Infleqtion, IQM Quantum Computers, Lawrence Berkeley National Laboratory’s Advanced Quantum Testbed and the United Kingdom’s National Physical Laboratory. (nvidianews.nvidia.com) The release fits into a wider Nvidia push to make its graphics processors part of the quantum stack, from the CUDA-Q software platform to the NVQLink interconnect and the Nvidia Accelerated Quantum Computing Research Center. Ising adds an open model layer to that strategy, aimed at speeding up the lab work between today’s noisy devices and larger fault-tolerant systems. (nvidia.com, developer.nvidia.com)