Skip to content

QF Solver benchmarks and reproducibility

QF Solver publishes bounded numerical and large-model evidence for selected finite-element routes.

This page is a public summary of reproducible measurements. It is not a claim that QF Solver has the same performance for arbitrary meshes, elements, materials, solver settings or hardware.

Large-model evidence

The strongest QF Solver 0.2.7 large-model evidence currently concerns structured TET4 linear-static models using selected PETSc/MPI routes.

Model size Recorded result Scope
1,029,000 DOF Two stable PETSc replays Structured TET4, recorded environment
3,000,000 DOF Two Silver replays plus bounded Gold Compute evidence Frozen PETSc/CG/GAMG route
5,012,640 DOF Two Silver replays 9,773,946 TET4, recorded PETSc/MPI environment
10M scale Bounded C3 evidence only Not presented as a general 10M performance claim

These results do not establish universal scalability.

5M Silver result

The 5M Silver campaign recorded:

  • 5,012,640 degrees of freedom;
  • 9,773,946 TET4 elements;
  • 1,243 solver iterations;
  • free residual near 9.85e-11;
  • equilibrium error near 1.39e-9;
  • energy error below 2e-14;
  • two recorded runtimes of approximately 4,428 s and 4,379 s.

The measurements describe this specific workload and environment. They are not a general performance guarantee.

Earlier scaling evidence

QF Solver also contains controlled scaling experiments for smaller structured TET4 models.

The historical matrix-free campaign includes completed iterative solves through 750,141 DOF and a bounded 1,029,000-DOF readiness attempt.

That campaign records:

  • model topology;
  • wall time;
  • memory consumption;
  • residuals;
  • sparse matrix information;
  • environment metadata;
  • deterministic replay information.

The historical campaign should not be confused with the later PETSc/MPI qualification evidence.

What is reproducible?

QF Solver attempts to preserve enough information to distinguish:

  • the model being solved;
  • element and material formulation;
  • solver backend;
  • convergence tolerance;
  • execution environment;
  • residual and equilibrium checks;
  • replay consistency;
  • resource-limited and solver-limited cases.

Failed or resource-limited attempts are not converted into successful results.

Reproduction and evidence

Detailed evidence for QF Solver 0.2.7 is available in:

Machine-readable qualification records are retained in the repository under:

qualification/0_2_7/

Interpretation

The published measurements should be used to answer:

Has this particular QF Solver route been exercised reproducibly at this scale?

They should not be used to claim:

QF Solver will solve every finite-element problem of this size.

Performance comparisons with other FEM solvers require equivalent models, elements, formulations, tolerances, solver configurations and hardware.