NVIDIA cuQuantum Adds Dynamic Gradients, DMRG, and Simulation Speedup
Table of Contents
- Dynamic Gradients in cuDensityMat
- NVIDIA Blackwell Kernel Optimizations in cuStateVec
- DMRG Primitives for Quantum Emulations in cuTensorNet
Dynamic Gradients in cuDensityMat
The latest update to cuDensityMat includes new APIs for calculating gradients of quantum state evolution. This allows for efficient backpropagation of quantum dynamics simulations with respect to optimizable Hamiltonian parameters, enabling QPU builders to train AI models for calibration, control, gate, and qubit design. The ability to compute gradients of target cost functions is essential for optimizing QPU layouts and drive pulses, reducing the timeline to useful quantum processors.
NVIDIA Blackwell Kernel Optimizations in cuStateVec
cuStateVec now features custom GPU kernels optimized for the latest NVIDIA GPU architecture, delivering performance improvements of 2-3x over NVIDIA Hopper systems. These optimizations ensure researchers can achieve the best performance on advanced NVIDIA hardware, especially for operations involving batching, expectation value calculations, and collapse operators.
DMRG Primitives for Quantum Emulations in cuTensorNet
The latest release of cuTensorNet introduces Matrix Product State Density Matrix Renormalization Group (MPS-DMRG) primitives, allowing developers and researchers to solve DMRG in quantum computing simulations. These primitives support solving DMRG through the MPS time-dependent variational principle (MPS-TDVP) algorithm, enabling modeling of longer-range interactions and larger Hilbert spaces with higher accuracy than existing methods.
To start using these new features, download cuQuantum and explore them in your quantum computing simulations or workflows. For questions, requests, or issues, visit the cuQuantum GitHub repository. Learn more about NVIDIA quantum computing here.