Track the frontier, from condensed matter to cosmology.
Search papers, labs, and topics across Lattice.
This paper introduces a scalable sample-based quantum diagonalization (SQD) solver, implemented in JAX for GPU acceleration, capable of subspace diagonalization up to $2^{30}$ dimensions. The solver addresses the computational bottleneck in finding ground states of general Pauli-sum Hamiltonians by providing an efficient eigensolver and a specialized matrix-vector operation function. Empirical results demonstrate the eigensolver's GPU memory efficiency and strong scaling performance in multi-GPU environments.
Sub-quadratic quantum diagonalization for Hamiltonians is now feasible on GPUs, reaching $2^{30}$-dimensional subspaces.
A scalable ground-state solver to be used in sample-based quantum diagonalization (SQD) is presented. The solver is built on JAX API, and is intended to be run on GPU platforms and possibly many-GPU high-performance computing facilities. It is capable of subspace diagonalization in up to $2^{30}$-dimensional subspace, with possible upgrades to enable even larger-scale operations. The two main components, a general-purpose eigensolver and an SQD-specific matrix-vector operation function, are described in detail. Efficiency of the eigensolver with respect to the GPU memory is demonstrated using a simple physics model. Furthermore, good strong scaling of the eigensolver in many-GPU environment is observed.