xTBloom: GPU Batching for High-Throughput GFN-xTB Calculations

Conformational searches and multiwindow free-energy sampling require large numbers of GFN-xTB energy and force evaluations across molecular conformations and sampling windows. Processing each system separately on a GPU makes it difficult to amortize memory management, host-device data transfer, and kernel-launch overhead. Here, we present xTBloom, a GPU batching library for GFN-xTB energy and force evaluations. xTBloom sorts evaluation queues by atomic orbital (AO) count before forming batches, accommodates size differences through a native ragged layout, maintains independent self-consistent charge (SCC) and failure states for each system, and selects execution paths according to data residency. We focus the quantitative evaluation on GFN2-xTB through cross-program comparisons on QM9 and OMol25, CPU-CUDA comparisons, finite-difference force checks, and controlled failure tests. On a near-size OMol25 workload, GPU execution achieved a 2.9-fold speedup over xTBloom's single-core CPU backend. A companion LAMMPS plugin integrates cross-window batching with periodic quantum mechanical/molecular mechanical (QM/MM) simulation. In a short synchronized test with 48 windows on one GPU, the complete coupled workflow achieved an aggregate speedup of approximately 14 over sequential execution of the same windows. These results demonstrate the throughput benefits of batched GFN-xTB evaluation for repeated electronic-structure calculations and multiwindow sampling.

Authors

Publication Details

Journal
arXiv (Cornell University)
Published
2026-11-18
DOI
https://doi.org/10.48550/arxiv.2611.24807
Primary Topic
Advanced Chemical Physics Studies
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

xTBloom: GPU Batching for High-Throughput GFN-xTB Calculations

Qiangqiang Gu, Jinzhe Zeng, Tong Zhu, Zitian Lu et al.
arXiv (Cornell University)
Advanced Chemical Physics Studies
preprint

xTBloom: GPU Batching for High-Throughput GFN-xTB Calculations

Qiangqiang Gu, Jinzhe Zeng, Tong Zhu, Zitian Lu, Shengrui Xu
preprint en

Abstract

Conformational searches and multiwindow free-energy sampling require large numbers of GFN-xTB energy and force evaluations across molecular conformations and sampling windows. Processing each system separately on a GPU makes it difficult to amortize memory management, host-device data transfer, and kernel-launch overhead. Here, we present xTBloom, a GPU batching library for GFN-xTB energy and force evaluations. xTBloom sorts evaluation queues by atomic orbital (AO) count before forming batches, accommodates size differences through a native ragged layout, maintains independent self-consistent charge (SCC) and failure states for each system, and selects execution paths according to data residency. We focus the quantitative evaluation on GFN2-xTB through cross-program comparisons on QM9 and OMol25, CPU-CUDA comparisons, finite-difference force checks, and controlled failure tests. On a near-size OMol25 workload, GPU execution achieved a 2.9-fold speedup over xTBloom's single-core CPU backend. A companion LAMMPS plugin integrates cross-window batching with periodic quantum mechanical/molecular mechanical (QM/MM) simulation. In a short synchronized test with 48 windows on one GPU, the complete coupled workflow achieved an aggregate speedup of approximately 14 over sequential execution of the same windows. These results demonstrate the throughput benefits of batched GFN-xTB evaluation for repeated electronic-structure calculations and multiwindow sampling.

arXiv (Cornell University)
Advanced Chemical Physics Studies
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