Simulation usage accounting
Simulation runs draw on a simulation allowance. RunMat measures that usage as the floating-point operations the run performs: meshing, assembly, solvers and result recovery. The same algorithm on the same model counts the same work however fast it runs, on however many cores or GPUs, so faster execution alone never changes the charge. Usage does not measure physical quantities, CAD dimensions, accuracy or elapsed time.
The included allowances are intended for simple brackets and everyday objects under free use, and larger design-and-simulation workflows under Professional. Usage depends on the physics, mesh, load cases, iterations and numerical algorithms; geometry alone does not determine the amount of measured work.
The unit
Usage is counted in GFLOP. One GFLOP is exactly 1,000,000,000 measured floating-point operations. GFLOP measures an amount of work; GFLOP/s measures speed and is not the unit of the allowance.
Larger amounts display on decimal scales of the same count:
| Display unit | Operations |
|---|---|
| GFLOP | 1,000,000,000 |
| TFLOP | 1,000,000,000,000 |
| PFLOP | 1,000,000,000,000,000 |
2,500 GFLOP and 2.5 TFLOP describe the same work. RunMat stores the exact integer count; the decimal display does not round what is charged.
What is counted
Each numerical step counts the operations of a defined reference algorithm. The count does not come from hardware instruction counters or from elapsed time.
| Operation | Measured operations |
|---|---|
| Real addition, subtraction, multiplication or division | 1 each |
| Fused multiply-add | 2 |
Real square root or transcendental function (sqrt, exp, log, pow, atan2, …) | 1 per evaluation |
| Complex arithmetic | Its real operations, for example 6 for a complex multiplication |
| Vector, SIMD or GPU arithmetic | The operations for each active element |
Single and double precision count the same. Parallel execution and instruction fusion add nothing. A different algorithm or a different number of iterations on the same model can do a different amount of work, and the count reflects it.
Metered steps follow reference schedules. For example, a real dot product of length n counts 2n − 1, and each conjugate-gradient iteration counts its matrix application, preconditioner application, dot products and vector updates. How each operation is counted lists the schedules for meshing predicates, sparse factorizations and substitutions, iterative solvers, result recovery and transient integration, and which schedules are still being written before accounting version 1 is published.
What is not counted
- OCCT CAD preparation and tessellation (excluded from accounting version
rsua-1). - In RunMat 0.7.0 (
rsua-1), two kinds of kernel whose pivoting or iteration has no approved reference schedule yet: sparse LU and Bunch–Kaufman factorizations and their substitutions, and the modal and buckling eigen kernels (dense symmetric eigen-decomposition, Lanczos and LOBPCG iteration, shift selection and eigenpair postprocessing). The solutions are still produced. The matrix applications, vector operations, preconditioners, LLT and LDLT solves and assembly around them are counted. - Integer arithmetic, comparisons, memory movement, communication and waiting. Some meshing work is therefore uncharged; RunMat assigns it no conversion weight.
- Input validation, field-shape checks and diagnostic formatting.
- Loading, viewing or copying a result that already exists.
- Ordinary math execution outside a simulation run.
Memory, CPU capacity and wall-time limits are separate from the usage meter. Cluster access and any cluster charges are separate entitlements.
Reuse, cancellations and retries
Reusing an existing output does not charge its construction again. New work that uses it is counted. Solving a new load case on an existing mesh counts the new assembly and solve, not the mesh. Applying an existing factorization to another right-hand side counts the substitutions, not the factorization.
Work is charged when a step completes:
- A step that is refused before it starts changes nothing and charges nothing.
- A step that fails before it produces a result charges nothing for that step. Steps completed earlier keep their charge.
- Cancelling after a step completes does not erase that step's charge.
- An iterative solve that converges early charges only the iterations it ran. Unused maximum iteration counts are never charged.
- An abandoned run expires 24 hours after its last usage report. It is charged for the completed work last durably reported, and the rest of its reservation returns to the allowance.
Reserved operations are not charged work. Recovery after an interruption must keep the identity of completed work so the same work is never charged twice; that mechanism remains a release requirement.
Estimates and spending caps
A pre-run estimate is not the charge. Adaptive meshing, nonlinear convergence and iterative solves can need more work than planned.
Every run has a fixed spending cap. The default cap is the available allowance plus eligible packs at admission. You can set a lower cap. Additional work needs admission before it would exceed the cap, and a grant renewal during the run does not raise that run's cap.
Usage receipts
Each run produces a usage receipt with:
- the accounting version;
- the run and attempt identities;
- completed work by category, such as meshing, assembly, factorization and solver iterations;
- the exact total operations;
- the settlement state.
Accounting versions stay published so past receipts can be read under the rules they were charged with. Example totals on RunMat pages come from actual runs.
Allowances and representative projects
The included simulation allowance and renewal period are published at https://runmat.com/pricing. The applicable Free Use grant or commercial agreement identifies which allowance applies. Eligibility verification for academic use is available at https://runmat.com/contact?type=academic.
Representative projects include measured receipts and enough information to reproduce their workloads. A full-house design workflow is a calibration target, not a guarantee about every house model or a measured release claim.
How each operation is counted
The per-algorithm schedules (meshing predicates, sparse LLT and LDLT, iterative solvers, structural result recovery and transient integration) are in How each operation is counted.