Authors: Chao Lu (Oak Ridge National Laboratory (ORNL)), Pooja Rao (NVIDIA Corporation), and Muralikrishnan Gopalakrishnan Meena and Kalyan Gottiparthi (Oak Ridge National Laboratory (ORNL))
Abstract: VQLS is a promising hybrid quantum-classical algorithm for solving linear systems Ax=b, but its practical deployment is hindered by the massive circuit execution overhead required when encoding arbitrary matrices via linear combination of unitaries (LCU) decomposition. Evaluating the VQLS cost function requires quadratic scaling requirement with the LCU terms, resulting in tens of thousands of circuits per iteration and millions of executions across a full optimization trajectory. To address this, we present a scalable, distributed implementation of VQLS that exploits the embarrassingly parallel structure of the gradient-free-based cost function using NVIDIA CUDA-Q conjunction with multi-node distribution for HPC system. Benchmarking on the Perlmutter supercomputer, we validated the VQLS algorithm with a sample tridiagonal Toeplitz and a canonical Hele-Shaw fluid flow problem. Furthermore, we demonstrate near ideal strong scaling and high efficiency weak scaling performance. We also used Nsight to benchmark the distributed framework to explore ideal concurrency.
Presentation: file