GPU acceleration of first-principles X-ray absorption spectra: an OpenACC port of XSpectra in Quantum ESPRESSO

XSpectra, the X-ray absorption module of Quantum ESPRESSO, computes K- and L-edge spectra from a Lanczos (Haydock) recursion on the self-consistent Kohn–Sham Hamiltonian of a supercell containing a core hole. In supercell calculations the recursion dominates the cost, but XSpectra had no GPU implementation, and in the GPU-enabled versions of Quantum ESPRESSO it stopped as soon as it applied the Hamiltonian. We present an OpenACC port of XSpectra for ultrasoft and projectoraugmented-wave (PAW) pseudopotentials, validated here with PAW datasets. The Lanczos vectors stay on the device for the whole recursion; the Hamiltonian and the inverse overlap operator S-1 are applied there, and only the recursion coefficients are returned to the host. The matrix that defines S-1 is built once per k point with cuBLAS. The CPU build gives the same results as before. We validate the port against the CPU code and benchmark it on a set of solids and molecules. One NVIDIA RTX PRO 5000 workstation GPU runs the XSpectra step 1.6–3.6× faster than a 64-core Xeon node for 64–162-atom cells, and 1.5–3.0× faster for the whole calculation including the SCF, which runs on the CPU; for the 162-atom α-SiO2 cell, the XSpectra step for the Si K edge is accelerated by 3.6×. For small cells with dense k-point grids the CPU node is faster. On the 162-atom cell, k-point parallelization over three RTX PRO 6000 GPUs is 2.94× faster than one GPU (11× faster than the CPU node), and the run time on different GPU models is ordered by their memory bandwidth rather than by their FP64 peak. The patches are available at https://github.com/YinanWang-iis/xspectra-gpu.

Authors

Institutions

Publication Details

Journal
ChemRxiv
Published
2026-11-18
DOI
https://doi.org/10.26434/chemrxiv.15012423/v1
Primary Topic
Parallel Computing and Optimization Techniques
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

GPU acceleration of first-principles X-ray absorption spectra: an OpenACC port of XSpectra in Quantum ESPRESSO

Teruyasu Mizoguchi, Yinan Wang
ChemRxiv
Parallel Computing and Optimization Techniques
preprint

GPU acceleration of first-principles X-ray absorption spectra: an OpenACC port of XSpectra in Quantum ESPRESSO

Teruyasu Mizoguchi, Yinan Wang
preprint en

Abstract

XSpectra, the X-ray absorption module of Quantum ESPRESSO, computes K- and L-edge spectra from a Lanczos (Haydock) recursion on the self-consistent Kohn–Sham Hamiltonian of a supercell containing a core hole. In supercell calculations the recursion dominates the cost, but XSpectra had no GPU implementation, and in the GPU-enabled versions of Quantum ESPRESSO it stopped as soon as it applied the Hamiltonian. We present an OpenACC port of XSpectra for ultrasoft and projectoraugmented-wave (PAW) pseudopotentials, validated here with PAW datasets. The Lanczos vectors stay on the device for the whole recursion; the Hamiltonian and the inverse overlap operator S-1 are applied there, and only the recursion coefficients are returned to the host. The matrix that defines S-1 is built once per k point with cuBLAS. The CPU build gives the same results as before. We validate the port against the CPU code and benchmark it on a set of solids and molecules. One NVIDIA RTX PRO 5000 workstation GPU runs the XSpectra step 1.6–3.6× faster than a 64-core Xeon node for 64–162-atom cells, and 1.5–3.0× faster for the whole calculation including the SCF, which runs on the CPU; for the 162-atom α-SiO2 cell, the XSpectra step for the Si K edge is accelerated by 3.6×. For small cells with dense k-point grids the CPU node is faster. On the 162-atom cell, k-point parallelization over three RTX PRO 6000 GPUs is 2.94× faster than one GPU (11× faster than the CPU node), and the run time on different GPU models is ordered by their memory bandwidth rather than by their FP64 peak. The patches are available at https://github.com/YinanWang-iis/xspectra-gpu.

ChemRxiv
Japan Graduate School of Education University (JP), Institute of Industrial Science, The University of Tokyo (JP), The University of Tokyo (JP)
Parallel Computing and Optimization Techniques
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon