Adaptive GPU execution planning for memory-limited Hamiltonian applications

Many quantum-simulation workflows, including propagation, filtering, Krylov, stochastic sampling, and observable measurement, are dominated by applications of Hamiltonians or other large operators, either to one state vector, ϕ=Ĥψ, or to several right-hand sides, Y=ĤX. These applications may be repeated many times, as in propagation and filtering, or may occur as one-shot observable, measurement, projection, or post-processing operations. The numerical algorithms used for these tasks are well established. The contribution of this work is an execution layer for these existing algorithms and operator evaluations when dense or effectively dense operator data exceed the memory of a single accelerator. The implementation first inspects the allocated hardware, including the number of GPUs (graphics processing units), available accelerator memory, and supported execution modes. It then enumerates candidate ways to apply the same mathematical operator: direct analytic action, seed-based generation, fused implicit kernels, on-demand block generation, parallel block execution, partial caching, full GPU caching, row-distributed caching across several GPUs, task parallelism, and hybrid row/task parallel execution. These candidates are ranked analytically using memory estimates, arithmetic-work estimates, cache requirements, and GPU-grouping constraints. The best analytically ranked candidates are then tested by a measured autotuner on the allocated hardware. Candidate comparisons use the same operator and input vectors and distinguish steady application time from calibration, compilation, cache construction, and tuning overhead. Plan selection minimizes expected total time for the declared number of operator uses. Validation against a fixed full-cache policy, direct GPU dense linear algebra, and a broader measured candidate search is used to quantify the benefit and overhead of adaptive planning. Logged trials may support learned ranking in future work, but no neural surrogate is used in the reported results.

Authors

Institutions

Publication Details

Journal
APL Computational Physics
Published
2026-11-19
DOI
https://doi.org/10.1063/5.0351591
Primary Topic
Parallel Computing and Optimization Techniques
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Adaptive GPU execution planning for memory-limited Hamiltonian applications

Uriel Shafir, Ronnie Kosloff
APL Computational Physics
Parallel Computing and Optimization Techniques
article

Adaptive GPU execution planning for memory-limited Hamiltonian applications

Uriel Shafir, Ronnie Kosloff
article en

Abstract

Many quantum-simulation workflows, including propagation, filtering, Krylov, stochastic sampling, and observable measurement, are dominated by applications of Hamiltonians or other large operators, either to one state vector, ϕ=Ĥψ, or to several right-hand sides, Y=ĤX. These applications may be repeated many times, as in propagation and filtering, or may occur as one-shot observable, measurement, projection, or post-processing operations. The numerical algorithms used for these tasks are well established. The contribution of this work is an execution layer for these existing algorithms and operator evaluations when dense or effectively dense operator data exceed the memory of a single accelerator. The implementation first inspects the allocated hardware, including the number of GPUs (graphics processing units), available accelerator memory, and supported execution modes. It then enumerates candidate ways to apply the same mathematical operator: direct analytic action, seed-based generation, fused implicit kernels, on-demand block generation, parallel block execution, partial caching, full GPU caching, row-distributed caching across several GPUs, task parallelism, and hybrid row/task parallel execution. These candidates are ranked analytically using memory estimates, arithmetic-work estimates, cache requirements, and GPU-grouping constraints. The best analytically ranked candidates are then tested by a measured autotuner on the allocated hardware. Candidate comparisons use the same operator and input vectors and distinguish steady application time from calibration, compilation, cache construction, and tuning overhead. Plan selection minimizes expected total time for the declared number of operator uses. Validation against a fixed full-cache policy, direct GPU dense linear algebra, and a broader measured candidate search is used to quantify the benefit and overhead of adaptive planning. Logged trials may support learned ranking in future work, but no neural surrogate is used in the reported results.

APL Computational PhysicsVol. 2(4)
Hebrew University of Jerusalem (IL)
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