← back to the showcase

Structural load-case families: 2.72× on the measured synthetic CalculiX study

A fixed ten-case structural job took 745.8–749.3 seconds with stock CalculiX and 272.5–277.1 seconds with the TAKT research build. The complete DAT and FRD results matched under the predeclared C-FEA-v1 contract below. Each observation covers the whole ten-STEP job, including output delivery.

The primary factor is 2.72×, the ratio of summed complete family times across three pairs. This is a result for one frozen workload and research build. Another model needs its own measurement.

The measured workload

Complete family times

W is one unmodified stock ccx process executing the complete ten-STEP job, with the required input staging and complete C-FEA-v1 output delivery. U is a fresh resident research process for the same job, including all run preparation, computation, checks, copies, output delivery and cleanup within the declared boundary. No computed state is supplied free from an earlier family, and no cost is spread over requests that were not executed.

Pair, each W then U W: one stock ten-STEP job, seconds U: complete research job, seconds W/U
1 745.762 273.331 2.73×
2 747.016 272.487 2.74×
3 749.302 277.121 2.70×

The ratio of summed times is 2,242.080 / 822.938 = 2.724481×, rounded to 2.72×. The median of the three individual ratios is 2.728425×; it is a different statistic.

Installation, compilation, preparation of the frozen fixture and independent output comparisons are outside both timing windows. Runtime preparation and full scientific output delivery within the declared job boundary are included. The recorded observer estimates were 2.0% for stock and 3.9% for the candidate; no time was deducted for these estimates or for the research build's checks. They are finite observed estimates, not confidence bounds or phase-share measurements.

All three pairs ran W then U on the same shared host. That order does not separate time drift from the comparison; no statistical confidence interval or future performance guarantee is claimed.

What matched

The declared contract is CANONICAL_EQUIVALENCE C-FEA-v1 of the complete DAT and FRD files:

All three complete family comparisons passed: six complete file comparisons across twelve stock/research delivery files. Each DAT was 1,257,540,810 bytes and each canonical FRD was 293,730,652 bytes. The full files were compared; fingerprints are additional records. Successful completion of all ten STEP records was checked. Diagnostic stdout, timing logs and solver diagnostic files are outside this scientific-output contract.

Measured scope

The result covers this model, ten prescribed CLOAD cases, fixed mesh/materials/boundary conditions, linear-static settings, pinned builds, one worker and the stated output contract. Other models, load locations, materials, constraints, family sizes, nonlinear/contact/thermal analyses, parallel solvers and other platforms have not been measured here.

The measured jobs completed under the stated address-space limit. A general whole-process memory bound and general failure-recovery guarantee have not been established. This is a research-result card, not a general CalculiX compatibility release or software-download announcement.

The wall-time factor depends on the share of native STEP time spent in factorization and on the remaining whole-job costs. The synthetic solid study above measured 2.72× on the fixed model with 497,664 nominal DOF across three pairs. A separate study of the public TEPS t11_CFRP_74 CFRP model, with 439,221 active equations and ten declared CLOAD cases, measured 1.47× in its first paired run: 88.0 seconds stock / 59.9 seconds research. Both paths used CalculiX 2.21, SPOOLES and one worker, and the complete DAT/FRD comparison passed C-FEA-v1. The TEPS priced study ended after that first family pair; no further priced family pairs or confidence bound are claimed. These are separate workload results; neither factor is a prediction for another model.

The public TEPS dataset records CC-BY 4.0 in its source metadata. The ten load variants were declared for this research study; this is not a record of a customer's production load cases.

Software and fixture provenance

The measured CalculiX 2.21 package records GNU GPL v2 for its main code. Packaging and individual components retain their separate licence notices. The first study uses a synthetic fixture; the separate public TEPS model has the CC BY 4.0 provenance stated above. This page supplies result text and record fingerprints, not an engine, a TAKT binary, source code or model files. Delivery and redistribution terms require their own specification. Project names identify the measured software; no endorsement is claimed.

Measurement record fingerprints

These SHA-256 values identify the retained internal records. The records themselves are not included on this page.

Record SHA-256
Three-pair measurement report dbba719af129c3728cabe082106c84dfd57007ab7011c6ab6ee0ff6c857dc44e
Independent source, contract and price review 154c07d50d7b7055df7bedf0df6855068b2d0e9cc4aada645b2d74da0f9db2c5
TEPS first-pair research report 85ff1f5b66c88fb99d8774b083e76f8e5c3cedf38e086aefdb8f4e4395a42d10

Describe your workload: engine and version, model size, case family, native runtime, required complete output and CPU allocation. The recorded factor does not promise a gain for another job.

Source text: README.md

Telegram